Fractal Image Coding: A Review

ARNAUD E. JACQUIN, MEMBER, IEEE
themselves. These objects, which arise from the mathematical theory of Iterated Sequences, were first labeled problem statement is the following. Given any original discrete mathematical “curiosities” or “monsters” by mathematiimage specified by an array of pixels, how can a computer concians in the beginning of the twetieth century who lacked struct a fractal imagMhe coded image-which is both visually the tools to properly analyze and understand them [12], close to the original one, and has a digital representation which requires fewer bits than the original image. [14], [15]. After falling into nearly complete oblivion for In this paper, we describe an approach to image coding based a while, they were “rediscovered” by the mathematical on a fractal theory of iterated contractive transformations defined research community in the 1970’s, thanks to the pioneering piecewise. The main characteristicsof this approach are that: i) it work of Mandelbrot who also coined their name [38]. It is relies on the assumption that image redundancy can be eftciently indisputable that this rediscovery was also triggered by the captured and exploited through piecewise serf-transformabilityon a block-wise basis, and ii) it approximates an original image by availability of computers and automatic graphic tools which a jhctal image, obtained from a finite number of iterations of made it possible for the first time to render and visualize an image transformation called a fractal code. We refer to this them as complex, beautiful, often realistic-looking objects approach as Fractal B k k Coding. or scenes [38], [47]. For the past ten years, fractals have The general coding-decoding system is based on the construction,for an original image to encode, of a fractal C O & also been part of a set of tools in a variety of fields in contractive image transformation for which the original image is Physics, where they are closely related to Chaos Theory an approximate jixed poin.-which, when applied iteratively on [19], [49]. They have recently emerged in various fields of any initial image at the decoder, produces a sequence of images Electrical Engineering, as attested by the contents of this which converges to a fractal approximation of the original. The Special Section. fractal code consists of a description of both an image partition and a contractive image tran$ormation defined as a list of parent Fractal-based techniques have been applied in several and child block transformations,each specified by a small set of areas of digital image processing, such as image segmentaquantized parameters. We describe the design of such a system tion [50], image analysis [51], [58], image synthesis and for the encoding of monochrome digital images at rates below 1 computer graphics 141, 171, [171, [221, 1231, [481, [541, blpixel, without any entropy coding of the parameters of thefractal and texture coding [37], {55]. Barnsley was the first to code. We also present novel ideas and extensionsfrom the work of a number of researchers which has appeared since the publication propose the notion of Fractal Image Compression, by which of the fractal block coding work of the author. real-life objects or images would be modeled by deterministic fractal objects-attractors of sets of two-dimensional affine transformations [2]-[4]. The mathematical theories I. INTRODUCTION of Iterated Function Systems (IFS) and Recurrent Iterated Deterministic Fractals have the intrinsic property of Function Systems [5], [6], along with the important Colhaving extremely high visual complexity while being very lage Theorem, constitute the broad foundations of fractal low in information content, as they can be described and image compression. However, these theories alone do not generated by simple recursive deterministic algorithms [38]. provide any constructive procedure for the “encoding” of They are mathematical objects with a high degree of a gray-tone image-as understood by the image coding redundancy in the sense that they are recursively made community-i.e., in an automated way. This particular task of transformed copies of either themselves or parts of can be performed by defining piecewise affine contractive transformations which make use of only the partial sevManuscript received July 8,1993. P a r t s of this paper are directly excerpted from a paper previously published by the author in the IEEE TRANSACTIONS transformability of images. ON IMAGE PROCESSING (vol. 1 , no. 1, Jan. 1992), entitled “Image coding In this paper, we are only concemed with fully autobased on a fractal theory of iterated contractive image transformations.” mated, robust, block-based fractal image coding schemes The work described in this paper was performed while the author was which can compress or encode any digital monochrome in the Mathematics Department, Georgia Institute of Technology, Atlanta, GA. image, and can therefore truly be called image compression The author is with AT&T Bell Laboratories, Murray Hill, NJ 07974schemes [20], [28], [42], [52]. The principles of Fractal 0636. IEEE Log Number 9212425. Block Coding described here were originally published in
Specifrc classes of Fractals can be used for Di-1 Image Compression or Coding. For such an application, the general
0018-9219/93$03.00 0 1993 IEEE
PROCEEDINGS OF THE IEEE, VOL. 81, NO. 10, OCTOBER 1993
1451

Under these conditions. 2) From (3). In a space of quantized discrete images. [35]. it is clear that the convergence of the sequence depends fundamentally on s being strictly smaller than one. [2]-[6]. In Section IV. they are meant to acquaint readers with block-based fractal image coding. including the one-dimensional case of contour encoding. and the specification of a class of discrete contractive image transformations defined blockwise from which fractal codes are selected. [36]. the sequence converges exactly to a stable image. present and analyze coding simulations. 10. Results are given for the coding of 512 x 512 8-b/pixel digital images such as “Lena. Sections II-V contain the mathematical principles as well as the description of a specific fractal image coding scheme. and say that porig is approximately self-transformable under T . and provided that T has a lower complexity than the original image. and let porigbe an original image that we want to encode. The result of this search is a discrete image transformation defined blockwise. A . The basic mathematical theory is succinctly presented in Section 11.” We also point to the analogies which exist with an image coding technique known as Vector Quantization (VQ). given a specific fractal block coder. VOL.which has a contractivity zero and which clearly satisfies (1) and (2). [241. NO. which as a result of its iterative construction. we describe the procedure for the encoding of any monochrome digital image. The Inverse Problem of Iterated Transformation Theory Let (M. d) and use of the contractivity of T. for any image po and any positive integer n: From (2) and (3). We call such a transformation T in F which satisfies (1) and (2) afractal code for porig. We denote by F the set of allowed transformations: a specific subset defined a priori of the space of all contractive transformations in (M. The requirements on 1452 the transfonnation T are formulated as follows: and d(Plorig7 T(Porig)) is as ‘‘close to zero” as possible. the selection of a distortion measure.By repeated application of the triangular inequality [131 in (M. The inverse problem of iterated transformation theory is the construction of a contractive image trakformation T. [38]. THEO~TICAL FOUNDATIONS In this section. [301-[331. We now describe the structure of the image transformations that can be used as fractal codes. For a broader introduction to data compression using fractal techniques. [161. This covers the construction of a partition of an image support. from 7. which is built so as to leave the original image approximately invariant and which we call a Fractal Code. (2) The scalar s is called the contractivity of T. and indicate the specificity of fractal block coding. 3) The constant transfonnation T = porig. d) to itself. [43]-[46]. 11. 1381. Section VI is a compilation of novel ideas and extensions which have been introduced by researchers following the publication of the initial work of the author.defined from the space (M. The procedure consists of the organized search of a “virtual codebook” obtained from a pool of domain blocks and a pool of block transformations. A block diagram of a very general fractal coding4ecoding system based on these principles is given in Fig. They are rooted in a general theory of iterated contractive transformations in metric spaces of images based on the work of Bamsley et al. for which porigis an approximate jixed point. T can be seen as a ~ossy’ image code for porig. and since s < 1. The purpose of this paper is twofold. we refer the reader to a paper by Beaumont [SI. a procedure which is typical of the construction of deterministic fractal objects [4]. d). we see that after a number of initial iterations the terms of any iterated sequence of the form: where po is some arbitrary initial image. OCTOBER 1993 . we present novel ideas and extensions from the work of a number of researchers [8]-[lo]. only a visually close approximation to porig can be recovered. will never be in F as we are only interested in transformations which can be unambiguously described with fewer bits than the original image-a requirement for compression. is said to befractal [4]. d) denote a metric space of digital images. T(Porig)). Notes: 1) The closeness of P ( p 0 ) to porigis conditioned by the distortion d(porig. ‘As opposed to “lossless”. This terminology is used because images produced by the above procedure are the result of the iterated application of a deterministic image transformation to an initial image. [57]. where d is a given metricdistortion measure.‘ I I [25]-[27]. 1. which have appeared since the publication of the original fractal image coding work by the author. we briefly present the underlying mathematical principles of fractal image coding based on a theory of contractive iterated image transformations. In Section V. cluster around the original image. and show how to evaluate bit rates. [41]. PROCEEDINGS OF THE IEEE. Section I11 addresses practical issues in the design of a digital image coder based on iterated transformation theory. 81. [39]. In Section VI. it is easy to show that. we address the decoding or reconstruction of an image from a fractal code.

B.: d ~ (~11 z s 7 fi1SI = ( (pio+i. denotes the restriction of the image sizes. B). The larger cells-of size B x B-are referred to as (range) parents. we call it the image block over R.(pOrig1D. JACQUIN: FRACTAL IMAGE CODING B. Structure and Construction of Fractal Codes for Images The realization that the redundancy in digital images can be captured through blockwise self-transformability leads us to focus our attention on a class of image transformations of the general form vp E M . A general fractal encodingdecoding system.image. = O<i<N (Tp)]R.) O<i<N = (5) where p = {&}O<i<N (6) denotes a nonoverlapping partition of the image support into N range cells-usually square blocks of possibly different 1 R. the smaller ones-of size B/2 x B/2-as (range) children. ~i is written as the composition of two transformations S. and iii) the specification of a finite class of contractive image transformations defined consistently with a partition. denotes an elemental block transformation from the domain cell Di to the range cell I&.)is a close approximation of the original range block porig] k . 2. This type of partition is closely related to quad-trees [561. For clarity. A partition constructed in this way is image-dependent. Small image b l o c k s 4 x 4 and below-are i) easy to analyze and to classify geometrically.(P) Fig. The symbol p p to the cell &.e. j o +j l2 O<aJ<B ) .x T. such that the transformed domain block T. and write p= O<i<N PIR.ig-a contractive image transformation of the form described above under which porigis approximately blockwise selftransformable-will be done by defining the elemental 7. Let p be an T. (7) simply to indicate that an image is the union of its restrictions to the partition cells. B) denote the square block of size B x B. = Ti o s i (8) where Si and T . ii) they allow a fast evaluation of interblock distances. [60]. and the peak 1453 . and ii) lead to higher compression ratios. Image Partitions The square support S of the original digital image porig is partitioned into nonoverlapping square range cells of two different sizes. with the bottom left comer at the intersection of image row io and image column j o . [591. 2. The selection of sizes and shapes for image cells depends on several factors. i. Distortion Measure We construct a distortion measure between digital images from an interblock distortion measure. and of a scheme for the quantization of their parameters. 1. and independently from one another. Other designs and their effects on compression rates and coded image quality will be discussed in Section VI. for every fractal block encoding of porig partition index i. Application of an image transformation T defined blockwise to an image p. and p i . and fi be an approximation of p. These will be specified in Section 111-C. DESIGN OF A FRACTAL BLOCK CODING SYSTEM The three main issues involved in the design of a fractal block coding system are: i) the (image-dependent) partitioning of an image. from a domain cell Di to the range cell R.jo+j + i . j the gray level of pixel ( i .. of the squared differences of pixel values. respectively. 111. Thus the amounts to finding.’s separately. 1 ’ 2 (9) We define the distance d between two images as the sum over the partition of all block distortions. . Let S(i0.j o . it allows the coder: i) to use larger blocks to take advantage of smoothly varying image areas. Ti(P1D. large b l o c k s 4 x 5 and above-i) allow a better exploitation of the redundancy in smooth image areas. Throughout this section the design choices are those of the author. iii) they are easy to encode accurately.I I Encoder Decoder Channel Fig. a transformation T. and T i ~i A. even when source images are diverse.The L2 or root-mean-square distortion between the image blocks p1s and /lis is defined as the square root of the sum over the cell S. Decisions about the splitting of a parent cell are made during the encoding of the image block over this cell (see Section V-C). We refer to the ith step as the encoding of the ith block of porigon the partition P.jo. The symbol T. and iv) they lead to a robust encoding system-one whose performance is steady. are the so-called geometric and massic parts of T. On the other hand. The application of such an image transformation T to an image p is illustrated in the diagram in Fig. The construction of a fractal code T for p0. thus forming a two-level square partition. Let p l s and fils denote their restrictions to the cell S(io. and ii) to use smaller blocks to capture detail in complex areas such as rugged boundaries and fine textures. A parent can be split into up to four nonoverlapping children. j ) . ii) the choice of a measure of distortion between two images.

1 ) Of block: (~4p)i. J ( j ) P Z ( ~ ) + I .j = &E-1-j. . Here. CONSTRUCTION OF FRACTAL CODES DIGITAL IMAGES This section addresses the implementation of a specific fractal block coder based on the theoretical coding procedure described in Section 1 1 . We list in the following a list of the eight canonical isometries of a square block. the results are given in Table 1.j iii) Contrast scaling by (Y + Ag- (14) E [0.. B )are the average values of four pixels in the domain block ( s ~ ) i . ( Table 1 LzContractivities of the Massic Transformations of Section IU-C (lo) Massic Transformation Absorption at go Luminance shift by Ag Contrast scaling by a Color reversal Isometries L2 Contractivity s=o C. VOL. . we describe the step-by-step construction of a fractal block code for any given original image.^ = pi.j- (16) so the contractivity of this particular block transformation is a. FOR PROCEEDIh'GS OF THE IEEE. j T . I) Geometric Part Si: It consists in the discrete form of the D : B spatial contraction operator S. (23) = go. Iv. NO. (15) The following transformations simply shufie pixels within a range block. the pixel values of the contracted image on the range block S ( i . through + B O 0 : In [25]. massic transformations allow to generate from a single block a whole family of geometrically related transformed blocks. As in Section III-B. we show that this operator has contractivity one. A few simple transformations defined in the space of discrete image blocks supported on a B x B range cell are listed below: i) Absorption at gray level go E ( ' U ~ ) O S ~ < G : (6P)ij 8) Rotation around center of block. (13) ii) Luminance shift by constant Ag E { & w ~ ) o < ~ < G : (Tp)i. and on the design choices made in Section III. In the simple case where D = 2B. j d . shuffling transformations can be constructed. + i . As an example. 10. = pj. This terminology is used in order to stress the fact that these transformations affect the pixel values of the block. through +goo: (L5p)i. A Class of Discrete Image Transformations In this section..= j~ (p + r( ji ) . we describe the class 3 of discrete contractive affine image transformations defined blockwise that the coder uses. (21) (12) 7) Rotation around center of block. OCTOBER 1993 2) Orthogonal reflection about mid-vertical axis ( j = (B .~. 1) Identity: LOP)^. which provides a pool in which matching blocks will be looked for during the encoding. 2) Massic Part Z:Massic transformations are those which process image blocks supported on a range cell R. cell R.j I(i)= id + 22 and J ( j ) = j d + 23. The L2 contractivities of the discrete block transformations described in this section can be easily computed. = S ( i r . ~ ( j )+ + l~ r ( i ) + 1 .1]: Some other. in a deterministic way-we call them isometries.j = Pi. ~ ( j ) + 1 ) / 4 .j = pE-l-i.j =p~-i-j.. .1) where the index functions I and J are defined by: (20) 6) Rotation around center of block. which maps image blocks from a domain cell D. D ) . to a range B).~--i-i.j.jo.j =Wa. + ~ z ( i ) .-. S denotes the cell S(io. J ( ~ ) s=l s = CY2 s=1 _ _ s=l 4) Orthogonal reflection about first diagonal (i = j ) of block 5) Orthogonal reflection about second diagonal ( i +j = + B (11) . j PE-I-. j E (0. (18) 1454 1 .B .E-l-i.signal-to-noiseratio (SNR) by S N R = 1Ol0g1o d:$yT2) where d r ( p ) denotes the dynamic range of p.1)/2) of block: (Llp)i. more complex. In effect. through -90": ( ~ 7 ~ l ) i= .l ) / 2 ) of block (L2p)i. B): (W)i.. (17) 3) Orthogonal reflection about mid-horizontal axis ( i = (B . = S ( z d .j. we derive below the contractivity of a contrast scaling by a. 81. j r . for all i .

made up of all image blocks which can be extracted from the original image. Shade blocks are not used as domain blocks. and De. C .. Overview o f the Procedure An original r x r digital image p. respectively. (27) B. 3The analogy with the notion of codebook used in vector quantization for image coding is discussed in Section V-D. D = D. but it can be trimmed and organized in order to make the search for an optimal domain block tractable. is given as input to the coder. 3. and which are larger than the range block R.e. " in tions T. (25) left comer at (0. of finding the block transformation Ti which minimizes the distortion between Ti o Si(p10. Using the thorough study of image block classification done by Ramamurthi and Gersho [53].The encoding of the range block p l ~ consists finding a "best pair" (D. Moreover. We can therefore remove them from the pool. which will be contracted to a block S i ( p 1 ~of ~ size ) Bx B. and midrange blocks. The total domain pool D is the union We call the product pool 2) x 7 a global pool or virtual codebook3. Let { & ) o ~ . Pools of Domain Blocks The maximal domain pool corresponding to a range block of size B x B can be thought of as all image blocks of size D x D ( D > B ) located anywhere in the image to encode. ( p l D i ) )is minimum.and mixed edge blocks Dme: De = D. since i) a shade block remains a shade block under any of the block transformations that we consider. It is important to note that the specification of the domain i (location and size) is equivalent to the description of cell D . 3. The description of a block transformation should be kept as simple as possible in order to obtain low 1455 . not to have any privileged orientation-"texture blocks" belong to this category. An initial domain pool D can be obtained by sliding a window of size D x D ( D = 2B is used) across the original image. quantized to 256 gray levels.) and p1Ri. it is easy to check that the nature of a domain block is preserved under application of any element of the transformation pool. It is assumed to be approximately isotropic.. The domain blocks in this pool D are then classified. (translation vector and contraction the spatial contraction S factor). at all times. It then moves from one position to the next by steps of either Sh pixels horizontally to the right. A shade block is "smooth" with no significant gradient. in such a way that it remains entirely inside the image support. which runs through the block. Encoding of the range block R. 0 The second part of the construction of 7-i consists of selecting the proper processing of the contracted domain block S. JACQUIN: FRACTAL IMAGE CODING Details about the classification algorithm can be found in P51. We can also formally define a pool of massic made of all discrete block transformatransformations 7.U Dme.O). The window is first located with its bottom ''The actual construction of this two-level partition is addressed in Section V-C. Ti) E D x 7. The steps can typically be chosen equal to B or B/2. the construction of a transformation ~i which maps onto this cell is broken into two distinct steps corresponding to the transformations Si and Ti.. and Dmr for shade blocks.I I A . The following two sections address a way to conduct Ti) by a priori disa directed search for a best pair (Di. ~ i . It is typically very large. An edge block presents a strong change of intensity across a curve-often a piece of object boundary. pixels vertically upwards. A midrange block has moderate gradient but no definite edge. d(p1~ ~o.. and ii) absorptions generate uniform shade blocks. s . < N denote the image partition made of range cells of two different sizes? Recall that an image transformation T has the form: T = n .i. (of size B x B). respectively. withTi=TioS. T. edge create three classes 'Ds...(ploi).. or 6. T O<i<N Given a range cell R. U De U Dm (26) where De is further split into two subclasses of simple edge blocks D. and by analyzing the range block to encode along with a set of candidate domain blocks.. 0 The construction of the spatial contraction S. amounts to selecting an image domain block plloi of size D x D. which is such that the distortion Fig. We can thus formally define a pool of domain blocks D. blocks. the resulting global pool D x 7. based on their perceptual geometric features.. [531. The encoding of a range block is illustrated symbolically in Fig.e. carding large subsets of the global pool. i. Pools of Transformation We now propose a procedure for the construction of and for the directed search of a transformation pool 7. The quantization of the parameters of the transformations is dictated by storage requirements. we define three types of blocks.

is computed so that p z . Image Reconstructionfrom a Fractal Code We use the original 512 x 512. The mapping of an image under the fractal code is done i is sequentially. a luminance shift. . We use massic transformations that are compositions of a contrast scaling. respectively. i. of one of the following. and the related construction of a candidate massic transformation T.e. . S . S. Finally. and Ag. where the function quant represents uniform quantization with 6 b in the range [0. Case 2a): p 1 ~ is . RESULTS + Agd. P I D J regions. with the initial image “peppers. . the transformation r applied to the current image block over the domain cell Di. )) are finally selected. . which is denoted Jil R. by a uniformly gray block. . is an edge. which generates this uniform block is simply the absorption at g.in the domain pool. In Fig. or the bright regions4epending on the dominant and S i q i -are the same. with initial image po. consists ) of a crude The analysis of the pair ( p l ~ . = quant (pL] R. The ~ ~ associated massic transformation domain block p 1 and Ti which minimize the distortion d(p1R. described in Section IV. . l ~ .. we restrict our attention to massic transformations that are compositions of a contrast scaling and a luminance shift..255].pT. pl D. of the form: T@. The optimal transformed domain block ~ . The segmented blocks are denoted pTi. until convergence to a stable decoded image is observed. populations of . . 4. = quant (pl R. After this segmentation is performed. We assume that image blocks of sufficiently small dimension can be segmented into two uniform regions.. Ag. a. = called ~ the The sequence of images {pn = P ( p ~ ) } r is reconstruction sequence for the code T .in the original image.” PROCEEDINGS OF THE IEEE. Note that the contractivities of the transformations can be obtained by multiplying together the contractivities of the elemental transformations which compose them. the function quant represents uniform quantization with 6 b in the range [-128. [25]. . ~ scanned. and mapped onto the range cell l&. our encoding procedure consists.) determines the selection of the transformation T~ used to encode p ] ~ . ( a % P . 10. We simply approximate p with gray level equal to the average gray level of p l ~ . same dynamic range It is then quantized to the nearest value in a set of four values smaller than one. ). a midrange or an edge block. a midrange. For midrange blocks. of the domain pool of the same Every element p l ~ . one dark. Secondly. OCTOBER 1993 .) + (28) Fig.p$. Case 2b): pi^.. ~ The analysis of a pair type as the block p 1 is (p1R.1271. Case I : p 1 R. is computed and quantized so that the gray levels of either the dark 1456 A.. o Si (p1D. 4 we show the first eight iterations of a decoding sequence corresponding to a fractal block code for the 512 x 512 “Lena” image. . (30) have the First. are done for every domain p 1 ~ . ql v. S). we compute the dynamic range of the segmented blocks as the gray level difference between the bright and dark regions. The analysis of the pair (p1R. ( p l D . p 1 ~ . and an isometry.) (29) where this time. First seven iterations of the “peppers-to-Lena” decoding sequence. one bright. separated by a transition region. The natural decoding scheme simply consists in iterating the fractal code T on any initial image PO. } o l ~which minimizes the distortion measure is selected. Case 2: p l ~ is . of the form T%(s~plD.‘ I I bit rates. For each cell index i. and S. g the one among the eight isomgtries { ~ ~ . where a. is computed so that the average gray levels of the range block and the scaled domain block are approximately the same. and S.p .based on the segmentation of the blocks p 1 ~ and computation of histograms of gray levels [20]. is a contrast scaling factor whidh is allowed to take values in a set of four values smaller than one. 81. is > called the matching block for p 1 R. 1 R.azpl D. according to the nature of the block. The transformation T. and a fractal code obtained through the encoding procedure.) = a%(s~plD. VOL. 8-bpp image “Lena” (Fig. is a shade block. .).P1DJ) = Ln. NO. p 1 D. For every range block p 1 ~ .T.

i=O Li(. The performance of the encoding system.o) i=O (34) is the partial sum of order N .i. Values of the SNR between the original “Lena” image and successive terms of this last reconstruction sequence are listed in Table 2. fixed set of coding specifications. Note that the remainder is the only term in the expression of p~ which depends on the initial image PO. i=O Li(. The finite sum N-1 Li(.0 26. &are displayed in Fig. 11) associated with the metric space ( M .1 dB of accuracy at the twelfth iteration. is displayed in Fig. and iii) the pool of transformations.vo (32) where L is a contractive linear image transformation. B. A coding simulation was carried out with an empirically chosen. T ( P ) = L ( P ) -t. a fractal approximation of the original. The decoded image. Because of the La-contractivity of 7. 11 . . 1457 .8 31. The first eight terms of this running sum-for the fractal code as in Fig. Under the heading “Encoding specifications. When the running sums of increasing order 1 1 2 3 4 5 6 7 8 9 10 S N R 10. i. The original image is the standard 512 x 512.we can see that they are affine [13] in the normed vector space (N. By looking at the form of the transformations in 3.o).” The image T L ~ ~ ~ ( present the encoder. The enhanced error image..2 30. in terms of bit rates and fidelity of the decoded image to the original. is 4Section V-C addresses the way bit rates are computed. The Nth iterate of the reconstruction sequence can be written as N-1 PN = i=O Li(uO)-k LN(PO)* (33) The above expression for . is the remainder of the expansion. 9. . . . and system performance-is given in Table 3.9 30. encoding procedure. It. the L2-energy of these terms decreases to zero. 10. it can be easily shown that IILN(Po)ll I sNlboIl (36) Table 2 Peak SNR Values Between the Original “Lena” and Successive Terms of a Reconstruction Sequence with an Initial Black Image. Convergence is obtained within 0. at ~ ) obtained for “Lena.4 20. and ii) the SNR’s between the original and the decoded image. and the residual term LN(PO) (35) Fig..1 17. i=O L i ( . It was initially split into four 256 x 256 subimages.e.” we define: i) the partition of the image support. is negligible for sufficiently large values of N .1 2 7 vo.1. 5 . equal to five times the pixelto-pixel difference between the original and the decoded image. d ) . 11. The term Li(v0) is the ith order term of the expansion. as the ratio of the number of bits necessary to describe the fractal code divided by the number of pixels in the image. Test Results Information about the specific fractal block coding system used for the encoding of “Lena”4esign specifications. 5. 8. 8-bpp “Lena” image shown in Fig.o). and vo is an image which is blockwise uniform.e. . We denote by the optimal fractal block code ~ ~ .It is interesting to consider another decoding scheme which illustrates the convergence of the iterative reconstruction procedure differently. which..1. By a result of Banach space theory [13]. and which indicates the image quality that can expected at the decoder (see (3)).UN is a series expansion of order N . It turns out that this required number of iterations remains approximately the same for different initial images. the images fade to black. 0 ) (37) of these terms are computed.8 24. since s < 1. is displayed in Fig. the reconstruction sequence corresponding to a uniformly black initial image is obtained. First eight iterations of the “black-to-Lena” decoding sequence (initial all-black image not shown). a code T has the form: VP E N . Under the heading “System performance” we list: i) the bit rate4.8 28.7 14. ii) JACQUIN: FRACTAL IMAGE CODING the pool of domain blocks. is shown in Fig. which were encoded independently from one another. No.

and the block is coded as four child blocks. VOL. = 4 b. midrange. according to the procedure presented in Section IV-C. indicating that optimal domain blocks are often found in the immediate vichity of the range blocks. twelve possible coding configurations of a parent block. with parent cells of size B x B . The configuration can thus be encoded with I. 12. therefore.. The following features of these histograms are of interest? i) the histograms of domain-to-range distances peak at zero.2 dB given in Table 3. = 4 shade (s). 8 bpp 0. centered near zero. which map to blocks of size B / 2 x B / 2 (child blocks). and iii) the quantization of the numerical parameters of these block transformations. with respect to its storage or transmission requirements. as is expected for the type of 5Some of these features have also been observed and exploited by other researchers in their own implementations of fractal block coders (see Section VI) 1458 C. 81.” It also appears that the quality of fractal codes relies very heavily on block classification. the reconstruction is free of edge degradation in the form of a “staircase effect. The coding of a range parent block is performed in the following way. L N.N. the initial parent transformation is discarded altogether. and Ie. Depending on these distortions. The image of the parent domain block under the best transformation found-the parent matching block-is then split into four children. Secondly. both of which are most likely to occur either with very large blocks or with blocks with either weak edges or strongly contrasted textures which are difficult to classify. No entropy coding of the transformation parameters was performed to obtain further compression.. midrange. from child transformations.e. For every child block such that the distortion is too high. which are represented symbolically in Fig. or rugged. 10. the coding of the parent block is or is not pursued to the next level. iii) the seven isometries of a square block listed in Section 111-C as ~1 through ~7-i. OCTOBER 1993 . First. mixed edge (me) Transformation Pool Shade block: Midrange block: Absorptions Contrast scalings Luminance shifts Edge block Contrast scalings Luminance shifts Isometries System Performance Test image Bit rate SNR (peak) SNR (signal energy to noise) “Lena”. The average bit rate is NpI. is essential to the evaluation of compression ratios or bit rates. were computed from the specifications of Table 3. If either three or four child transformations are needed. The decoded image has an overall very good fidelity to the original. However. NeIe bPP (38) NpBz + + + %hen dealing with two-level square image partitions. ii) the nature of the block transformations used. depending on the nature of the range block being coded (shade. are slightly “smoothed out” by the encoding process...I. Three distinct types of block transformations are used. ii) the histograms of luminance shifts are approximately Laplacian. 6. are very accurately preserved. 512 x 512. PROCEEDINGS OF THE IEEE. Computation of Bit Rates The complete description of an image transformation T. It is due to the use of square range blocks. or edge). we first compute the information in bits needed to represent a block transformation of each type.06 bpp 31. N . we distinguish parent transformations.4 x 4 (child) memoryless block coding method used here. a child transformation is constructed and is “appended” to the parent transformation. and child cells of size B / 2 x B / 2 . The important factors are: i) the description of the image partition. contrasted contours. simple edge ( s e ) . whether they are smooth. I. 6 and child transformations for T L are and 7. 16 x 16 (parent) 8 x 8 (child) bh = 4. and Ne denote the total numbers-parents and children-of (range) shade. Sharp. Image “textures” are well preserved although some finely textured areas.. the nontrivial rotations and reflections-are rather seldom used. In order to compute a bit rate given complete specifications of a fractal block coding system. which map to image blocks of size B x B (parent blocks). and edge blocks in the image. Histograms of the parameters of the parent ~ plotted ~ ~in Figs.Table 3 Design Specifications and Performance of a Fractal Block C d i g System EncadiPartition Subimage size: Type of partition: Range blocks: Domain Pool Domain blocks: Domain block shifts: Classification: Specifications 256 two-level hierarchical square 8 x 8 @arent). There are. Some artifactual “blockiness” is visible. and analysis. we consider the case of a two-level square partition. and are shown in Table 4. These quantities. midrange (m). Most of the artifacts visible in reconstructed images are rooted either in wrong block classification or in inaccurate block analysis.4 dB 24. Let N. and distortion measures are computed between these and the child blocks in the original image. NO. denoted respectively as I. a pool of parent transformations6 is searched for a best transformation. such as the turban around the hat.

:. 6 Luminance shifts t 0..11 0. 6.. 70 . The fractal code consists of the description of the imJACQUIN: FRACTAL IMAGE CODING age partition... ~ ~ I .. 0 1 2 3 * aSi Fig. . 7 .. 0 0 . ? ~ ...' I . Domain to range distances contrast s c a l i i 0. . .05 . '. 5 5 0 . 7 0. 7..05 0 A ..(' . ! ' 2 .... along with that of the image transformation defined as the ordered list of range block bansformations: { ( 7 ~ . c 0 ..7 6 * 0 ~ 3 4 5 7 Fig. 1 2 .. 5 < N} (39) 1459 . Statictics of a fractal code (part I: parameters of parent transformations).05 0 n. .. 4 0' 5 6 7 0 . 1 .. 8 5 1 Ui Luminance shift5 0 0 -256 -128 ...-> ' 0 .. . Statictics of a fractal code (part II: parameters of child transformations). Ia . . 7 5 2 ' 1 ... -64 0 64 0 128 256 &i5 o~~L.~ 1 ) .. 4 50 . * . 22 ) . 0.. where Np = (r/B)' is the total number of parent blocks in the image..2. .. 8 1 : 0 ~ -128 -64 0 64 128 AX AY d 0 .

at the decoder. and also partly on the complexity of the image to encode.Analogies with Vector Quantization Fractal block coding is in some sense related to the image block coding technique known as vector quantization (VQ) [ 11. ’ are four child transformations which are possibly appended to the parent transformation ~ i Each . [ 181. 81. = 19 bits (p) I. of these block transformations are from the pools of tranformations defined in Section IV-C. rate. 8 bpp. Decoded “Lena” image obtained after eight iterations of men=on an initial black square. = 18 bits (c) Edge D. T : . I. = 6 bits Midrange parent: 4 +4= 8 child: 5 + 5 = 10 2 6 a. The encoding does not rely on the existence of any PROCEEDINGS OF THE IEEE. hence the term “virtual codebook. something which can be remedied by improving the searching strategy for matching blocks (see Section VI). Our coding system is therefore variable1460 a coding system. we give a comparative list of features shared by both methods. we found that for given specifications of D. 10. The structure of a fractal code is illustrated in Fig. which the arrows indicate block transformations from domain block to range block. T:. = 16 bits (p) I . In Table 5. 9.” Table 4 Information in Bits for the Representation of Block Transformations Block Type Shade Parameters Information in Bits 6 s z D. +4 = 8 child: 5 + 5 = 10 parent: 4 2 6 3 Fig. Encoded “Lena” image mena(porig) As.’. = 21 bits (c) where T-:>’. OCTOBER 1993 .” The particular system described in Sections IV and V is also plagued by high encoding complexity. Fig. However. VOL. 0. Original 512 x 512 “Lena” image. NO. 13. Enhanced error image for “Lena. 8. We use VQ terminology. (domain) I.Fig. Also. Fig. The bit rates that we obtain depend partly on the system specifications. [21]. 10. and not needed. I . It is important to stress the following points: Domain image blocks are not present.6 bpp. for the construction of the fractal code.” This “codebook” is used only during the encoding phase. 11. 31. the selection of image partitions and allowed classes of block transformations could be made in compliance with a maximum bit rate which would be computed for a “worst encoding case. images with comparable block statistics (in terms of percentages of blocks of different types) are encoded with comparable bit rates. and are specified by a set of quantized parameters. T-:~’.4 dB. and relate it to the specific terminology introduced in this paper. [53].

[411. pool of block transformations. it is characteristic of fractalbased image generation techniques. we recall that the block transformations we used for each block type are affine and of the general form: 1461 . [26]. Square blocks.[16]. For (child) range blocks of size 5 x 5. Direct reconstruction by look-up table I Encoding of range blocks Fractal code: list of block transformations. we present a number of extensions which have been proposed by various researchers [8]-[IO]. Trimming of domain pool. iii) the comparison and possible merging of fractal block coding with more traditional blockbased image coding techniques. VI. 1 2 4 1 . 13.1 5 7 1 . Block matching Decoding Image code: list of addresses of blocks in the codebook. and optimization of the parameters defining these transformations. and segmentation algorithms used in this paper becomes apparent. A .[391. the crudeness of the block classification.These extensions broadly address the following issues: i) the influence of the type of image partition. Off-line transmission of codebook. [301-[331. blocks.[361.[431-[46]. Coding configurations for a parent range block. analysis. These features are what make fractal block coding very different from vector quantization. Image degradation is of roughly the same type for both types of block-coding methods. As for the iterative reconstruction of an image. EXTENSIONS In this Section. although sharp edges are very well preserved by our system. ii) the reduction of the complexity of the encoding process-the search for best matching blocks in the image to encode. Codebook design. It is interesting to note that the largest block size about which strong image degradation begins to occur is the same in both systems. since the original fractal image block coding work was published JACQUIN: FRACTAL IMAGE CODING by the author in 1989-1990 [25]. No transmission of domain Codebook Training set of images. Iterative reconstruction. We now present some very recent novel ideas and conclusions from these studies below. Optimaliiy o f Transformation Parameters From Section IV-C. Structure of a fractal image block code. “extemal codebook” but on the redundancy present in the original image itself. Ui -EFig. VQ Fractal Black Coding Image Partition Square blocks. processed domain blocks extracted from original image. and iv) the extension of the theory to the three-dimensionalcase for fractal block coding of sequences of images.‘ I 1 parent no child ( 1 configuration) 1 parent 1 child ( 4 configurationr) 1 parent 2 children ( 6 configrrationr) no parent 4 children ( I configuration) Fig. 12. Different block sizes possible. The “virtual codebook” is in other words “image-adaptive’’ in the sense that it is provided by the original image itself. Global pool (virtual codebook) Contracted. Different block sizes possible.

along with a value of amax equal to 1. It was observed by several researchers [lo]. are optimally derived by using a least squares criterion. j < B. and the overall search complexity is greatly reduced.. the upper bound on a0 for all the block transformations which constitute the fractal code. Other Block Transformations and Search Complexiry Reduction In [43]. this latter term was a priori restricted to be a flat luminance shift over the range block R. PROCEEDINGS OF THE IEEE. C . is motivated by the desire to reduce even more drastically the search for domain blocks-the main factor for long encoding times with block-based fractal image coding. on the other hand. In [44]. [26]. and a 3 . When combined. these two isometries specify an isometry which maps domain block to a range block with the same orientation. but with translation terms of the form 3 Ag = k=l akak + 3 bkjk k=l +to (44) k=l where { a l . respectively. this constant term can be viewed as a firstorder. i. therefore propose to first GramSchmidt orthogonalize the fixed-basis vectors { A I . [ a ] .e. as parent and child blocks. j < B (43) B. Jacobs et al. Other Partitions. be of the form subspace of R B x Bi. . j ) = i A 3 ( i . with no perceivable degradation of image quality but with a great reduction in search complexity. The authors use an innerproduct space formulation of the problem of optimally deriving the three parameters. The operator (a0J o S) is the linear part of the transformation. Candidate matching blocks are written as a linear combination of the processed domain block J ( Spl D)-the image-dependent part of the matching.I I where the indices corresponding to the particular range and domain block considered have been dropped for simplicity.A z .A z . for 0 5 i . j= ) a$ + ali + b z j 2 + blj + coij + t o . (42) The constant term Ag can be geometrically visualized as a tilted plane in the three-dimensional space formed by the plane of the image with a third dimension for gray levels.. In [24].A s } are the fixed basis vectors consisting of the three matrices A l ( i . The authors report a significant increase in image quality by allowing these higher order shifts. the authors define smooth 1462 which are third-order polynomials in i and j . as well as the best choice for a . A3). In [16]. ap. rely on an even “richer” pool of affine block transformations as in (a). 81. j )= 1 A 2 ( i . In [24]. then Gram-Schmidt orthogonalize each candidate domain block J ( S p l ~to ) the fixed basis. This scheme is reported to be only very slightly suboptimal. 10. blocks to be those image blocks which are approximated sufficiently well by an optimal quadratic polynomial in i and j . Oien et al.A z . the authors report a reduction in encoding complexity by an order of magnitude. for 0 5 i . as compared with the zero-order approximation case. By allowing optimal quadratic polynomial approximation of smooth blocks. [41] that one could do away with the seven isometries of a square block listed in Section In-C. Since these authors allow very few domain blocks to be searched. . It was also reported by a number of authors [lo]. OCTOBER 1993 . In [25]. which corresponds to the image-adaptive part of the matching process. VOL. all nontrivial rotations and reflections. for which no search is necessary. the optimal coefficients could be easily computed without having to invert matrices. Furthermore. Lepsby. respectively. If the basis vectors happened to form an orthogonal set. This results in having to recompute only the optimal coefficient a0 for each new domain block considered.e. and presented in [39]-[41]. as ~1 through LT-i-e. whereas the term Ag is the translation term. Q Z . (dien.5. amaxcan be chosen to be greater than one. A subsequent technique proposed by MONOand Dudbridge. which we summarize here. In [44]. compared with the full search of eight isometries. show that it is not necessary to impose that every block transformation be strictly contractive. and Ramstad suggested that the translation term Ag be chosen in a three-dimensional . entirely defined by the three coefficients Q I . two-dimensional polynomial in the row and column indexes. a3) are real coefficients. they. This scheme finds. Equivalently. for a pair of domawrange blocks. and the three fixed basis vectors { A I . Other Block Shapes We described in Sections 111-A and V-C an encoding procedure for a two-level square partion of an image into 8x 8 and 4x 4 blocks. referred to. Fisher et al. say the orientation which puts the brightest block quadrant in the upper left comer and second brightest in the upper right one. Jacobs and Fisher conducted a thorough study to determine the optimal number of bits for the uniform quantization of a0 and Ag. and convergence to a stable decoded image will still hold. with no cross products. the row and column indices of pixels in a range block.. the isometries which will give a specific block orientation. of the form ~ ( i . and { A I . in effect decoupling the optimization of the scalar coefficients for the constant and linear parts in (40) and (41). j ) = j . The study in [24] leads to the optimal choice of uniform quantization of a0 and Ag with 5 and 7 b. the percentage of image blocks which can be encoded by only a constant part increases dramatically. NO.A3}. arrive at the same complexity reduction-without throwing away the isometries altogether-through the use of a block orientation scheme. now defined by seven parameters instead of one. These constant terms.

as well as that of combining fractal block coding with other more traditional block coding techniques. Somewhat different results are reported in [30]. Of course. as adaptive DCT. Quadtree-based image partitions should be used. and 4 x 4. and to specify a class of contractive block transformations with quantized parameters. but increased coded image quality.16 x 16. given constraints of maximum complexity or bit rates.‘ I [24]. whether the same coded image quality as in the still image case could be obtained at reasonably low bit rates. fractal coding performs better than VQ. The preliminary design issues are to select an adaptive image partition made of nonoverlapping range cells. searching for matching domain blocks of size 12 x 12 x 12. with perhaps a slight advantage for VQ coding of texture block with an LBG codebook trained on texture blocks. but not as well. it is useful to both eliminate the initial strict contractivity requirement. The encoding of an original image then consists of capturing the self-transformability of the original image by searching a global transformation pool for a transformation defined blockwise-a fractal code-under which the image is approximately invariant. as affine block transformations may not efficiently exploit temporal correlation. or a best codeword picked in some image-independent codebook. In [8]. little change in compression. [33]. Fractal block coding has also been compared to other more traditional block coding techniques such as adaptive DCT coding [24]. The possible use of block shapes other than square was also evaluated. along with the number of publications which describe other implementations of fractal block coders for the encoding of images are exciting. Although the results are very preliminary. i. in a complex manner. Even though a complete reconciliation of the approaches taken by different authors may seem provides difficult at this point.e. report that their best implementation of a fractal block coding system performs comparably. However. compression. in [24]. to select an interblock distortion measure. JACQUIN: FRACTAL IMAGE CODING VII.8 x 8. For 1463 . It appeared that for sharp-edge blocks. The work of other researchers who either extended the theory. He uses a combination of three-dimensional fractal coding with a DPCM [29] prediction technique for the average luminance of range blocks. it is interesting to mention an extention of block-based fractal image coding to the encoding of digital video sequences by extending the theory to the three-dimensional case @I. [45]. In [45]. CONCLUSION We have described the design of digital image coding systems referred to as Fractal Block Coders which are based on a theory of iterated contractive image transformations. Decoding complexity remains fairly low and stable for various system designs. Specifically. The interest that the publication of the initial work has generated. *Sharp edges and textures. local “flatness” in images could be better exploited7. they efficiently exploit the flat areas present in typical images. It is unclear 7This has also been observed by other researchers in different image coding contexts [ll]. In [24] Jacobs et al. the body Of insight to the design of a fractal block coder for images. more traditional block coding techniques such as transform coding or vector quantization. (YO = 0.with domain blocks always twice the respective size of the range blocks they attempt to match. In [24]. This can be seen from (40) where these techniques can all be modeled as complex constant tranformations with no linear part. D. the following has been observed. or optimized the original system design to improve the initial results of the author was then presented. [57] that by allowing more than one level of recursion in the partition.. these authors report that perceptually annoying block effects occur. [59]. Jacobs et al. Attempts to determine which coding techniques are preferable for which classes of blocks are reported in [30]. [a]. Different formulations of the block-matching problem of finding the best combination of domain and block transformation for a given range block can lead to efficient least squares optimizations of the parameters of an optimal block transformation. thereby considerably reducing the complexity of the initial full-search scheme.Fractal Block Coding and Other Block-Based Image Coding Techniques Fractal coding is compatible with other. The relationship between vector quantization and fractal block coding was analyzed. A specific block-base fractal image coding system was presented as well as encodingldecodingresults. remains a fairly novel image coding technique. designed with the Linde-Buze-Gray (LBG) algorithm [34] were trained on image blocks of the same type as the blocks to encode. where we compared coding results obtained by using either fractal coding for texture and edge blocks or a simple memoryless VQ scheme. Fractal block coding. While keeping the requirement that the block transformations are affine. where codebooks of different sizes. the degree to which some of these issues are more important than others will have influence on what is considered an “optimal” fractal coding system. Ag could be either a DCT approximation of the block. Beaumont encodes slabs of four frames at a time. the authors report that it might be advantageous to encode the “complex” regions* of an image with fractal block coding and the “smooth” regions with a block-DCT encoder. as described in this paper. Domain blocks with sides tilted at a 45’ angle with respect to the sides of a square (or rectangular) image were used by Jacobs et al. For example. but the results are very similar for texture blocks. use a quadtree partirioning method to generate a partition which divides the image into nonoverlapping range blocks of sizes 32 x 32. as well as enlarge the pool of block transformations by allowing polynomial translation terms. These authors report that including such domain blocks results in higher encoding times. The choices in the design a block-based fractal coding system affect all of the following: visual quality of coded images. encoding complexity and speed.

Englewood Cliffs. vol.-optimal attractor image coding with fast decoder convergence. vol. CA. Boss. UK: Cambridge Univ. G . Dimensions non EntiLres et Applications. V. Gonzalez and P. pp. PAMI-6. Foumier. -. A. and D. Gleick. pp.’’ in Proc. CA: Freeman. K. ICASSP-92. 1053-1055. S. 1991. Berlin. M. “Generalized fractal transforms: Complexity issues. Nut. “Recurrent iterated function systems. 3-3 1. pp. E. Dudbridge. H-0. S. van Schooneveld. E... “An algorithm for vector quantizer design. Haskell. New York Academic Press. -. vol. 1001.-Apr. 81. 1990. M. and D. “Fractal image coding techniques and contraction operators. Musgrave (Co-chairs). M. N.” Signal Process. -. New York Wiley. F. 3. Carpenter. Roy. D. and M. Falconer. S.” Delft Univ. The Norwegian Institute of Technology. 1975-1977. 1991. Norway. J. Germany: Springer-Verlag. Acad. Ervin. B. NO. course notes. “Fractal approximation of image blocks. Lepsay. vol. Netravali and B. Mar. Machine Intell. A. Li. 4. IEEE Computer Soc. 251-263.” Electron.D. Trondheim. W. ICASSP-93. J. New York Springer-Verlag. New York: Plenum. 1989. Gersho and R. pp. and L. M. Ramstad. N. ACM. K. 1993. J. no. M. 19g1. dissertation. 1988. S. Mandelbrot. aien. T. KOMPRESJON-89 (Oslo. D. 11. Keane.. D. London. pp. D. [38].. Lepsay. New York: Vicking.” Proc. 1993. Y. vol.. Buzo. Chaos. no. Georgia Tech. M. Commun. “An imer product space approach to image coding by contractive transformations. The Beauty ofFractals. V. pp.” in Proc. Bamsley and A. 131-140. vol. 1. Fractals. Bamsley. “L.. 1989. 1991. 1988. Sandin. “Vector quantization. San Diego. pp. D. S. REFERENCES H. Addison-Wesley 1977. “Computer rendering of stochastic models. 371-384. Kauffman.. Monro and F. Monro. OCTOBER 1993 . A. The Geometry of Fractal Sets. “Texture coding using a pyramid decomposition. 1992. W. H a r t and F. Trondheim. 1982. 18-30. IEEE. no. The Norwegian Institute of Technology. “Fractal modeling in 3D computer graphics and imaging. F. Digital Image Processing. Jiirgens. Lepsay. III. “Fractal-based descriptions of natural scenes. fractal image coding performs quite comparably with more traditional image coding techniques. 1986. F. H. P. and R. H-0. Bedford. 1991.” IEEE Trans. A. G. Jayant and P. Oct. “Classified vector quantization of images. “Image coding based on a fractal theory of iterated con[27] -. 485-488. tractive image transformations. “Adaptive transform coding via quadtree-based variable blocksize DCT.. Technol. M. “Harnessing chaos for image synthesis. Hardin. J. Ramstad. P. San Francisco. Y. J. TIT. “Fractal coding of monochrome images. Gray. 25. M. F. J. Digital Coding of Waveforms:Principles and Applications to Speech and video. M. 1984. [24] E. 1982. Mar.” presented at NOBIM-92.. pp. Making of a New Science.” in Proc. R. aien. 243-275.. 337-340. “Reducing the complexity of a fractal-based image coder. vol. F. Fractals Everywhere. Image Process. Digital Image Compression Techniques. Mathematical Foundations and Applications. Fussell. 1993. Dekking.” AT&T Tech. vol. I all its esoteric origins.” Delft Univ. 1992. Jacobs. Vector Quantization. H. Some might find this disappointing.” Preprint. “Fractal signal modelling for source coding. 1985. pp. SPIE. and T. Y. M. and R. E. Dordrecht. S. pp. M. R.” Constructive Approximation. Gray. 1980. Rabbani and P.” in Proc. Saupe. Boss. Commun. 1991. New York Wiley. 1993. VOL. Hart. K. Malassenet. 69.” in Proc. Press. “A fractal-based image sequence compression scheme.” Proc. C. vol. Feb. vol. 1987. “A fractal theory of iterated Markov operators with applications to digital image coding. Jacquin. New York IEEE Press. Fisher. Cherbit. and T. We think that one main point of this work is that fractal block coding shows that there are ways to capture and exploit image redundancy in images. H. 84-95. pp. 1854-1857.D. “Image data comuression: A review. 6. ICASSP-93. Saupe. C. Dekking. W. 1990. no. F. 1985. Fractal Geometry. and T. “Application of recurrent iterated function systems to images. 91-100. Jacquin. 9. M.” IEEE Trans. 1988. Bamsley. Norway. 1992.J. Bamsley and S. The property of piecewise self-similarity and its capture through the construction of contractive image transformations which leave original images approximately invariant provides a new scheme for the exploitation of image redundancy for image compression-this property is what makes fractal image coding work. “Fractal surface models for communications about terrain. PROCEEDINGS OF THE IEEE. “Attractor image compression-fast algorithms and comparisons to related techniques. May 1992. 2225-2228. Paris. Sept. A. E. 1987. T. Hardin. Technol.” Ph.” in Proc. which are not used by more traditional image coding techniques. London. Trondheim.. Data Compression Conf. Pattern Anal. C. A399. D. 1. .” SPIE Visual Comun. COM-28. May 1989. Norway). LaurenGot and A. Jacquin. Pentland. 10. France: Masson. Forchheimer.. “Ray tracing deterministic 3-D fractals. “A novel fractal block-coding technique for digital images. “Image data compression using fractal techniques. Sollid. “Iterated function systems and the global construction of fractals. 1989. Graph. A. A. pp. Gray. “Solution of an inverse problem for fractals and other sets. 1992.. vol. “Fractal block coding of images. and R. 1992. D.” IEEE ASSP Mag.Jacobs.” Ph. Nov. Monro and F. vol. 28.D dissertation. T. The Science of Fractal Images. The Netherlands: Kluwer. 1992.” in Proc. SIGGRAPH ’88. N. “An approach to fractal coding of onedimensional signals. Peitgen and P. this despite the existence of a restricted class of complexlooking fractal objects which can be fully described by a very small set of parameters [4]. Breeuwer. Digital Pictures: Representation and Compression. 1984. ICASSP-91. Sci. G. pp. NJ: Prentice-Hall.” Ph.3. 1988. H-0.’’ BT Technol. Bellingham. 1986. COM-34. Memo. F. Rep. C. Jacquin. M. 122-131. M. Lundheim. Richter. pp. 1992. Schwartz. “Hybrid image block coders incorporating fractal coding and vector quantization.. 1992. Press. Peitgen.” IEEE Trans. [23] J. E. Jan. vol. H... Lett. Jain. and L. Malassenet. Lancaster. T. E. A.. Soc. 83. Intemal Rep. Dunford and J. with a robust classification scheme. MA. Press. Jones. The Fractal Geometry o f Nature. Naval Oceans Systems Center. vol. 1987. Chaos and Fractals. 93-109. Noll. Chen. “Image compression: A study of the iterated transform method. Novak. Image Process.” Ph. B. 1990. pp. 23. no. ICASSP-89. Bamsley. A. 1989. G. Beaumont. New York Springer-Verlag. Berlin: Springer. F. “Iterated transform image compression. Demko. and R. Dec.” in SIGGRAPH ’91.” in Proc. Elton. Fisher. pp. pp. 29. G. vol. [25] A. vol. Ramstad. M. M. Vector Quantization and Signal Compression. 2773-2776.” Commun. and L. WA: SPIE Optical Eng.. Bamsley. 1989.” presented at EUSIPCO-92.. pp. 0ien. F. dissertation. M. J. G. TR-1408.” IEEE Trans. Abut. Reuter. [28] A.” Comput. Reading. “Hybrid image compression combining transform coding and attractor coding. Peitgen and D. Ed. M. 1984. Ramamurthi and A. pp.” Proc. Linear Operators. Intemal Rep. Bedford.” in Proc. vol. H. -.D dissertation. Jan. A.” Proc. L. M. 1986. aien. Keane. Wintz. 1992. Dudbridge. Norway. Linde. 845. and D. Atlanta. and J.” NOSC Tech. Apr. [26] -. “Attractor image compression with a fast non-iterative decoding algorithm. ICASSP-90. 1993. The Norwegian Institute of Technology. 254-261. 353-356. 1966. Gersho. D. F. J.

IEEE) was born in Reims. France. Atlanta. 2040-2060. [55] R. “Fractal image models and object detection. 188-260. pp. dissertation. he has been a Member of the Technical Staff at AT&T Bell Laboratories.” which was published in the January 1992 issue ON IMAGE PROCESSING.” Delft Univ. vol. 1987. June 1984. Since 1990. Stein. [57] D.-Apr.. II. pp. Sum. computer vision. in 1964. Mar. He received the Diplijme d’Ing6nieur degree in electrical engineering from Eole Sup6rieure d’Electricit6. Technol. His research interests are in the areas of image and video compression.S.” in Proc.. pp. Signal Process. 1987. C. [59] J. 16. in 1986.. degree in mathematics in 1989. 1992.’’ ACM Comput. where he has been engaged in research on IOW bit rate video coding for teleconferencing/videotelephony. 40. Data Compression Conf. Zakhor. Vaisey and A. He received the M. [60]-. no. Atlanta.. degree in electrical engineering in 1987.” SPIE Visual Commun. Hodges-Reuter. and computer graphics. fractals. vol. Samet. [56] H. “Rendering and magnification of fractals using iterated function systems. 2. 1051-1054. Aug. p. Gif-sur-Yvette. D r . “Fractal coding of monochrome images. NJ. Jacquin is the recipient of the IEEE Signal Processing Society’s 1993 Senior Award in the area of Multidimensional Signal Processing. of the IEEE TRANSACTIONS JACQUIN: FRACTAL IMAGE CODING 1465 . France. both from the Georgia Institute of Technology. “Fractal approximation of images. Murray Hill.. no. Technol. ICASSP-87.D. Gersho. Intemal Rep. “Image compression with variable block size segmentation. van Schooneveld. Georgia Inst. and the Ph. 1987. 1993. 845. [58] M.D. “Variable block-size image coding. 451.[54] L.. 1991.” Ph. Arnaud Jacquin (Member. for his paper “Image coding based on a fractal theory of iterated contractive image transformations.” IEEE Trans. Rinaldo and A. Image Process. 8. in the Signal Processing Research Department. Dec.” in Proc. vol. “The quad-tree and related hierarchical data structures.

Sign up to vote on this title
UsefulNot useful