You are on page 1of 6

2018 4th Iranian Conference on Signal Processing and Intelligent Systems (ICSPIS)

Leaf Classification for Plant Recognition with Deep


Transfer Learning
Ali Beikmohammadi Karim Faez
Department of Electrical Engineering Department of Electrical Engineering
Amirkabir University of Technology (Tehran Polytechnic) Amirkabir University of Technology (Tehran Polytechnic)
Tehran, Iran Tehran, Iran
dr.abm@aut.ac.ir kfaez@aut.ac.ir

Abstract—Plant recognition systems that developed by In computer vision, despite the many efforts that have been
computer vision researchers, help botanists in faster recognition conducted, plant recognition seems as a challenging and
and detection of unknown plant species. Until now, multiple unsolved issue [1, 3, 9-12]. This is because a plant in nature has
studies focused on the process or algorithms that maximize use of a very similar shape and color representation that is shown in
botanical datasets for plants prediction modeling, but this Fig. 1. In addition to plants leaf morphological variations such
method depends on leaf characteristics which can be changed as changes in the size, texture, shape, venation, and so on, has
with botanical data and different feature extraction techniques. been added to this challenging issue.
On the other hand, recently, due to the popularity and successful
implementation of deep learning-based methods in various areas Characteristics such as shape [13-15], texture [16-18] and
such as image classification, object detection, and speech venation [10, 19] are features that generally used for
recognition, the researchers directed from traditional feature- recognition leaves of different species. Two basic types of
based methods to deep learning. In this research, one more feature representation methods for describing leaf images are
efficient method presented that use transfer learning to recognize hand-crafted features [10, 13, 16] and deep learning features [7,
plant for leaf classification, which first uses a pre-trained deep 20-22]. In practice, the design of hand-crafted features depends
neural network model for learning useful leaf characteristics on the ability of computer vision experts in encoding
directly from the input data representation. Then use a logistic morphological characteristics that predefined by botanists [20].
regression classifier for leaf classification. It is seen that transfer However, the deep learning features are able to be
learning from a large dataset to limited botanical dataset in plant
automatically learned based on the advantage of deep learning
recognition task is well done. The proposed method is evaluated
on two well-known botanical datasets, i.e., Flavia with 32 classes
algorithms. Representation based on learning (especially deep
and Leafsnap with 184 classes, and has succeeded in achieving an learning), introduce end-to-end concept by using a trainable
accuracy of 99.6% and 90.54%, respectively. The results show feature extractor and in continue of that introduce trainable
that despite the large change in the number of classes in these two classifier [23, 24]. Deep learning is a new and modern method
datasets, the proposed method, have a good performance and for image processing and data analysis that show high potential
show the better result than methods based on the hand-crafted and promising results. Deep learning has been successful in
feature and other methods based on the deep learning in terms of various fields and recently entered in the field of agronomy and
memory and precision. food products [8]. Therefore, leaf recognition based on deep
learning methods has attracted more attention.
Keywords—leaf classification; plant recognition; deep learning;
transfer learning; MobileNet In any way, the training of a new deep learning model from
the beginning requires a lot of data, cost, more computing
resources and hours, in some cases, days of training. On the
I. INTRODUCTION other hand, in real applications, collecting and labeling large
One of the most important issues in agronomy for plant amounts of data in a particular domain is costly and time-
researches and agronomy products is plant recognition and consuming issues, and may even be impossible. Therefore, in
classification (about 50 thousand species) [1-3]. A good many cases, the idea of collecting a large number of data in a
understanding of plants for help to recognize new and rare particular domain may not be acceptable [25, 26]. This
plant species is essential for improving pharmaceutical problem poses a challenge and makes it difficult to use deep
industries, the eco-system balance, agronomy utilization and learning models. The researchers believe, the reason that
robustness [4]. Botanists use variations on the leaf humans in their lives were able to learn thousands of species on
characteristics as a comparative tool for their study about plants just a few examples is that they were able to obtain this ability
[4, 5]. On the other hand, due to the progress of science and by collecting knowledge over time and transferring it to new
genetic topics, many of the compound plants cannot recognize objects for learning them [27]. In fact, the researchers believe
by ordinary people or even expert people in this field [6]. For that recognizing the objects in the past helps in learning new
this reason, tried to use computer vision or machine learning objects by their similarity and relation to new objects. Based on
techniques to solve this issue [7, 8]. this idea, some studies show that deep learning models that
trained for a classification task can be used for new

978-1-7281-1194-0/18/$31.00 ©2018 IEEE


classification task [28-30]. Thus, trained Convolutional Neural II. RELATED WORK
Network (CNN) models on a dataset, after fine-tune can be This section discusses state-of-the-art methods for plant
used for a new task even in a different domain [31-33]. This classification using hand-crafted and deep learning
concept is known as transfer learning. The use of transfer representation as a feature extractor.
learning to adapt existing neural models for translation, speech
synthesis, and visual classification tasks in the areas of object
recognition, image classification, and many other domains A. Hand-crafted Features
have been successful [34]. There are many hand-crafted features for the images of
plant leaf that can be divided into three general categories,
Basically, transfer learning is possible in two ways: shape [1, 2, 13-15, 48], texture [16-18, 49], venation [10, 19].
• Keeping the pre-trained network and updating the Most studies use shape recognition techniques for modeling
weights based on the new training dataset. and represent the shape of leaf contours. In one of the first
papers in this field, Neto et al. [13] introduced discriminant
• Using the pre-trained neural network as a feature analysis and Elliptical Fourier for detection of various plant
extractor and then representing with the help of a species based on their leaf shape. Subsequently, two modeling
generic classifier, such as a logistic regression for approaches based on invariant-moments and centroid-radii
classification [35]. were proposed [48]. The Shape Context (SC) and the
Histogram of Oriented Gradients (HOG) are also used to create
In fact, this can be said to be a convenient way to bootstrap
a leaf shape descriptor [14, 15]. Kumar et al. [1] proposed
your learning process. The second method has been successful
Histogram of Curvature over Scale (HoCS) for leaf analysis.
for many detection and classification tasks [36-38]. Our
Hall et al. [2] proposed a method using Hand Crafted Shape
proposed method also belongs to the second category for the
(HCS) and HoCS for analyzing the leaves.
classification of plants. In fact, the main idea is to test the latest
CNN-based models and architectures and combine it as a The texture is another important field of study in plant
feature extractor with a logistic regression classifier. We have recognition and it is used to describe the leaf surface based on
reviewed deep models scores that previously presented such as the distribution of pixels over a region. One of the oldest
AlexNet [39], GoogleNet [40], ResNet [41], Xception [42], studies [49] has used multi-scale fractal dimension for plant
VGG16 [43], VGG19 [43], ResNet50 [44], InceptionV3 [45], classifications. Subsequently, Cope et al. [18] use the analysis
InceptionResNetV2 [46] and MobileNet [47]. Based on of leaf margins to classify species and proposed a method by
experiments, MobileNet is selected as a feature extractor for using Gabor co-occurrences in plant texture classification.
classifying the leaves. Then a logistic regression as a generic Naresh and Nagendraswamy [16] corrected the conventional
classifier for leaf classification has been used. The remaining Local Binary Patterns (LBP). They take into account the
sections of the paper are as follows: related work, proposed structural relationship between neighbor pixels and use it to
method, experimentations and results, and conclusions are replace the hard threshold method that used in the conventional
presented in sections II, III, IV and V respectively. LBP. Tang et al. [17] introduced a new texture extraction
method based on a combination of Gray Level Co-Occurrence
Matrix (GLCM) and LBP to classify tea leaves.
The venation of the plant leaf is also widely used as a leaf
characteristic. Charters et al. [10] designed a new descriptor
named EAGLE, which consists of five patches of the sample
that is tuned to capture and extract spatial relationships
between the local regions of venation. They showed that the
combination of EAGLE and SURF would improve the
discriminative ability of feature representation. Larese et al.
[19] recognize different types of legumes based on the
venation. They first segmented the vein pattern using the Hit or
Miss Transform (UHMT) and then used LEAF GUI criteria to
extract a set of venation features. In addition, a significant
amount of research combines the mentioned features to show
plants leaves. For example, attempts have been made to
combine the shape and texture of the leaves in [11, 50] and
then continue with the addition of leaf color features [9].
Although the above methods have been successful, it
should be noted that the performance of these solutions
depends heavily on the set of selected features that are
themselves dependent on the task or dataset. As a result, in
studies on species recognition, the search for an effective
subset of the features for showing leaf samples has been
disturbed. On the other hand, this feature engineering process is
Fig. 1. An example of the Flavia leaf images dataset [55]. It can be seen that
all plant species have almost the same shape and color.
difficult and requires expertise and experience in that particular

22
field. Due to these limitations, research has led to a deep In this situation, transfer learning will be beneficial and
learning-based approach. solves this problem by using a pre-trained deep representation
as a source architecture for creating a new architecture [54].
B. Deep Learning Features Actually, with transfer learning, we use an existing trained
Recently, several plant leaf recognition methods based on model and adapt it to our own problem. We rely on the features
deep learning have been proposed [7, 20-22, 51]. Liu et al. [21] and concepts learned during the basic model training. We use
use a traditional CNN to extract features and then use a support the features learned from ImageNet datasets by a convolutional
vector machine (SVM) to classify leaf images. Guillermo et al. DNN with discard final classification layer, and combine it
[22] first segmented the vein pattern using UHMT to create with a logistic regression as a generic classifier that predicts
vein binary segmented images, then they substitute original our class labels in the new domain. We easily reuse the pre-
input images with these binary segmented images to train a trained features. To do this, we can train all the weights of the
CNN. Lee et al. [7] proposed the DeepPlant network to modified network or train only the weights of the new
recognize plant leaf images. In addition, they use predictive layer (that is, the logistic regression classifier). We
deconvolutional network [52] to gain insight into the selected use the second method in this paper.
features of the CNN model. Lee et al. [20] in continue with In fact, first, by the pre-end layer of MobileNet
their previous work, proposed a two-stream convolutional architecture, whose weights were pre-trained on the ImageNet
neural network (TwoCNN). The TwoCNN is a two-stream dataset and imagined to be frozen for our purpose, for each 224
feature learning that trained on the whole and patch of images x 224 pixels input image, a 50176-component feature vector is
respectively. Although TwoCNN can take discriminative extracted. Then, the set of these feature vectors is used to train
information on a variety of scales (both the whole and the a logistic regression classifier. The training parameters for
patches of images), the training process requires a more logistic regression classifier are: Penalty L2, Tolerance 1e-4,
complex sample set. Because you have to provide both whole Liblinear Solver, Maximum Iteration 100, Verbose 0, and OVR
images and segmented images for this network. Hu et al. [51] Method.
tried to use a combination of multi-scale features with a
convolution neural network called MSF-CNN to classify plant In this paper, after checking the latest CNN-based models
leaves on multiple scales. In this way, the multi-scale features and architectures, MobileNet [47] selected as a source
trained by the MSF-CNN are fused step-by-step and do not architecture to solve the plant recognition problem with the
require multiple learning branches. help of a generic classifier. The MobileNet is trained on the
ImageNet dataset and received an RGB image with a size of
However, deep learning methods also have some 224 x 224 pixels and classified into its corresponding class.
limitations. If the number of available images is less than the This architecture consists of fourteen convolutional layers and
number of images required for state-of-the-art based models thirteen Depthwise convolutional layers that are one among
such as GoogleNet network, we will be challenged to train the each other, a fully connected layer and a softmax classifier
network. As mentioned above, Transfer Learning is a useful [47]. This architecture includes 4.2 million parameters. As
technique in this situation. If your task is similar to a problem shown in the block diagram in fig. 2, We have used this model
that network has already been trained for it, you can use the as a feature extractor and followed by that use a Logistic
transfer learning to modify network to use it in your problem regression classifier to leaf classification. The proposed work
with a small set of labeled images. In fact, in transfer learning offers an interesting combination of deep learning and generic
method, can solve data limitation problem with the use of a classifier, which causes an increase in leaf classification
pre-trained networks as a source architecture for train target efficiency. In addition, with a 17MB of storage space for
models on small datasets [53]. MobileNet weights, this can be done for mobile application.
Fortunately, the winner models of the ImageNet Large
Scale Visual Recognition Challenge (ILSVRC), such as IV. EXPERIMENTATIONS AND RESULTS
AlexNet [39], GoogleNet [40], ResNet [41], Xception [42], This section discusses the conduction of the experiment, the
VGG16 [43], VGG19 [43], ResNet50 [44], InceptionV3 [45], training process and the test results of the proposed method.
InceptionResNetV2 [46] and MobileNet [47], are publicly Using the MobileNet model as a feature extractor, for each
available as pre-trained networks. These networks can be used input image 50176 features extracted, which set of this called
for transfer learning. In this paper, we are going to introduce a feature vector. Further, with the help of the image label and this
transfer learning-based method for the task of plant feature vector, the logistic regression classifier was trained to
classification. predict the target class. The proposed method is tested on two
known botanical datasets Flavia [55] and Leafsnap [1] with 32
III. PROPOSED METHOD and 184 classes.
Transfer learning is very useful with the use of deep CNNs The input of the MobileNet trained on the ImageNet dataset
to train a model with a limited size dataset. This is because is an RGB 224 x 224 image, which we will preserve the same
CNNs on a small dataset surely get overfitting. To overcome for our inputs, and resize them to this size and consider them as
this problem, the number of training data should be increased. network inputs. An explanation of these datasets and a
But providing a large amount of training data is very difficult comparative analysis is provided in the following sections.
and costly.

23
TABLE I. COMPARE PERFORMANCE ON THE FLAVIA LEAF DATASET
[55]. DL REPRESENTS A DEEP LEARNING-BASED APPROACH.
Year DL Feature Classifier Accuracy
Rate (%)
- Y MobileNet Logistic 99.6
Regression
2017 Y DeepPlant [20] MLP 99.4
2015 N HCF [2] RF 91.2
2015 N HCF-ScaleRobust [2] RF 89.8
2015 N Combine [2] Sum rule 97.3
(RF)
2015 N Morphological,FD,SDF ANN 96.0
[56]
2015 N HOG (Multi-scale Gaussian 94.7
Fig. 2. Proposed system overview, first row represents the source window) [20] SVM
architecture and second row shows the target architecture. 2016 N Modified LBP [16] NN 97.6

TABLE II. COMPARES THE PERFORMANCE ON THE LEAFSNAP LEAF


A. Evaluation on Flavia leaf dataset DATASET [1]. DL SHOWS A DEEP LEARNING-BASED APPROACH.

Wu et al. [55] introduced one of the famous datasets in leaf Year DL Method Rank-1
recognition field that is called Flavia. This dataset contains identification rate
1907 leaf images from 32 different species. The size of each - Y MobileNet 90.54
image is 1,200 x 1600 pixels that samples of them are shown in 2018 Y MSF-CNN [51] 85.28
Fig. 1. We such as [20], consider all leaf samples from each 2018 Y Plain-CNN [51] 83.35
2017 Y DeepPlant [20] 54.77
class, and 10 samples selected randomly for testing. We have
considered the whole leaf image as an input and did not use the
image patches. Finally, we compared our results with the other This dataset currently covers 184 tree species from the
state-of-the-art methods that previously presented in terms of northeastern United States. In this work, only the field images
average accuracy. are used because they are closer to real images for the
In Table 1, we can see that MobileNet model using logistic evaluation of the proposed method. Since the Leafsnap
regression classifier with the best average of classification training, validation and testing set are not available for
accuracy with 99.6%, and this method outperforms other state- researchers individually, so in this paper like [51], experiments
of-the-art methods that using the shape and statistics features are performed by randomly dividing the dataset into two parts.
[2, 56], texture features [16, 57] or deep learning features [20]. So that half will be used for training and the other half will be
Based on these empirical results, it is possible to find out that used for testing. This process is repeated 5 times and the
the learned features with the unsupervised method are good for average cumulative match curve (CMC) [58, 59] is computed
plants leaf representation. The superiority of the proposed as the final result. Euclidean distance is used as a similarity
approach compared with DeepPlant [20] does not limit to criterion between a pair of plant leaf images.
precision, but since the MobileNet model has fewer The performance on the Leafsnap dataset is shown in Fig.
parameters, less depth, and less memory requirements for the 3. Hu et al. [51] have been able to retrain DeepPlant [20] on
storage of weights than the AlexNet model, the proposed Leafsnap and found that DeepPlant to be very difficult
method will be faster and can be used on a mobile application. convergent. As shown in Fig. 3 and Table 2, while rank-1
The reason for all these excellences can be found in the identification rates of the MSF-CNN [51] and DeepPlant are
extraordinary architecture of MobileNet. 85.28% and 54.77%, respectively, the rank-1 identification rate
of proposed method gained 90.54%. In addition, it can be seen
B. Evaluation on Leafsnap leaf dataset that the proposed method has obtained a better CMC curve
For more researches in leaf recognition field, Kumar et al. than MSF-CNN and DeepPlant.
[1] published the Leafsnap dataset that includes images of The results show that the proposed method has lower
leaves taken from two different sources and also segmentation attenuation in its performance in comparison with other
produced automatically for them, as follows are: methods that have proposed their results on these two datasets
• 23147 Laboratory photo, consisting of high-quality and because of that, we will ensure its performance stability
images of compact leaves, from the Smithsonian against different datasets. The reason for this loss of
series. These images are provided in several samples performance is due to the presence of more classes and an
for any plant in versions with controlled background increase in between class similarities due to the proximity of
and foreground lighting. different classes to each other.

• 7719 Field images, consisting of ordinary outdoor It should be noted that since none of the previous work in
images provided with a mobile phone (often by this area has mentioned the processing time, so a fair
iPhone). These images contain a variety of blur, noise, comparison cannot be made from this view. However, due to
lighting patterns, shadows and so on. the use of the MobileNet architecture, it can be ensured that our

24
proposed method has a remarkable advantage in terms of REFERENCES
processing time than other deep learning based methods. [1] N. Kumar et al., “Leafsnap: A computer vision system for automatic
plant species identification,” in Computer vision–ECCV 2012: Springer,
2012, pp. 502-516.
V. CONCLUSION
[2] D. Hall, C. McCool, F. Dayoub, N. Sunderhauf, and B. Upcroft,
In this paper, we find with various experiments that the “Evaluation of features for leaf classification in challenging conditions,”
combination of the pre-trained MobileNet deep neural network in Applications of Computer Vision (WACV), 2015 IEEE Winter
architecture with the logistic regression classifier show good Conference on, 2015, pp. 797-804: IEEE.
accuracy for plant leaf classification. MobileNet trained for the [3] C. Kalyoncu and Ö. Toygar, “Geometric leaf classification,” Computer
Vision and Image Understanding, vol. 133, pp. 102-109, 2015.
new task as a feature extractor machine and logistic regression
classifier trained on the target dataset. [4] J. S. Cope, D. Corney, J. Y. Clark, P. Remagnino, and P. Wilkin, “Plant
species identification using digital morphometrics: A review,” Expert
It has been shown that with the help of the transfer learning Systems with Applications, vol. 39, no. 8, pp. 7562-7573, 2012.
technique, we can save in time and computation resource and [5] J. Clarke et al., “Venation pattern analysis of leaf images,” in
be successful in learning a new task by limited training dataset. International Symposium on Visual Computing, 2006, pp. 427-436:
Springer.
In addition, the proposed method works directly with RGB
[6] K. P. Ferentinos, “Deep learning models for plant disease detection and
images and thus eliminate the needs of pre-processing and diagnosis,” Computers and Electronics in Agriculture, vol. 145, pp. 311-
hand-craft feature extraction. The efficiency of the proposed 318, 2018.
method is evaluated on two botanical datasets, Flavia and [7] S. H. Lee, C. S. Chan, P. Wilkin, and P. Remagnino, “Deep-plant: Plant
Leafsnap, and precision of 99.6% and 90.54% achieved identification with convolutional neural networks,” in Image Processing
respectively. The comparative analysis confirms that the (ICIP), 2015 IEEE International Conference on, 2015, pp. 452-456:
proposed method outperforms the similar state-of-the-art IEEE.
methods for plant recognition using transfer learning. [8] A. Kamilaris and F. X. Prenafeta-Boldú, “Deep learning in agriculture:
A survey,” Computers and Electronics in Agriculture, vol. 147, pp. 70-
90, 2018.
[9] A. Kadir, L. E. Nugroho, A. Susanto, and P. I. Santosa, “Leaf
classification using shape, color, and texture features,” arXiv preprint
arXiv:1401.4447, 2013.
[10] J. Charters, Z. Wang, Z. Chi, A. C. Tsoi, and D. D. Feng, “Eagle: a
novel descriptor for identifying plant species using leaf lamina vascular
features,” in Multimedia and Expo Workshops (ICMEW), 2014 IEEE
International Conference on, 2014, pp. 1-6: IEEE.
[11] T. Beghin, J. S. Cope, P. Remagnino, and S. Barman, “Shape and texture
based plant leaf classification,” in International Conference on
Advanced Concepts for Intelligent Vision Systems, 2010, pp. 345-353:
Springer.
[12] J. S. Cope, P. Remagnino, S. Barman, and P. Wilkin, “The extraction of
venation from leaf images by evolved vein classifiers and ant colony
algorithms,” in International Conference on Advanced Concepts for
Intelligent Vision Systems, 2010, pp. 135-144: Springer.
[13] J. C. Neto, G. E. Meyer, D. D. Jones, and A. K. Samal, “Plant species
identification using Elliptic Fourier leaf shape analysis,” Computers and
electronics in agriculture, vol. 50, no. 2, pp. 121-134, 2006.
[14] S. Mouine, I. Yahiaoui, and A. Verroust-Blondet, “Advanced shape
context for plant species identification using leaf image retrieval,” in
Proceedings of the 2nd ACM international conference on multimedia
retrieval, 2012, p. 49: ACM.
[15] X.-Y. Xiao, R. Hu, S.-W. Zhang, and X.-F. Wang, “HOG-based
approach for leaf classification,” in Advanced intelligent computing
theories and applications. with aspects of artificial intelligence: Springer,
2010, pp. 149-155.
[16] Y. Naresh and H. Nagendraswamy, “Classification of medicinal plants:
an approach using modified LBP with symbolic representation,”
Neurocomputing, vol. 173, pp. 1789-1797, 2016.
[17] Z. Tang, Y. Su, M. J. Er, F. Qi, L. Zhang, and J. Zhou, “A local binary
pattern based texture descriptors for classification of tea leaves,”
Neurocomputing, vol. 168, pp. 1011-1023, 2015.
[18] J. S. Cope, P. Remagnino, S. Barman, and P. Wilkin, “Plant texture
classification using gabor co-occurrences,” in International Symposium
on Visual Computing, 2010, pp. 669-677: Springer.
[19] M. G. Larese, R. Namías, R. M. Craviotto, M. R. Arango, C. Gallo, and
Fig. 3. shows the CMC curve and rank-1 identification rates (R1) on P. M. Granitto, “Automatic classification of legumes using leaf vein
Leafsnap leaf datasets [1]. (up) the proposed method, (down) introduced image features,” Pattern Recognition, vol. 47, no. 1, pp. 158-168, 2014.
methods in [20, 51]. [20] S. H. Lee, C. S. Chan, S. J. Mayo, and P. Remagnino, “How deep
learning extracts and learns leaf features for plant classification,” Pattern
Recognition, vol. 71, pp. 1-13, 2017.

25
[21] Z. Liu et al., “Hybrid Deep Learning for Plant Leaves Classification,” in [42] F. Chollet, “Xception: Deep learning with depthwise separable
International Conference on Intelligent Computing, 2015, pp. 115-123: convolutions,” arXiv preprint, p. 1610.02357, 2017.
Springer. [43] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
[22] G. L. Grinblat, L. C. Uzal, M. G. Larese, and P. M. Granitto, “Deep large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
learning for plant identification using vein morphological patterns,” [44] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
Computers and Electronics in Agriculture, vol. 127, pp. 418-424, 2016. recognition,” in Proceedings of the IEEE conference on computer vision
[23] F. Zhu, L. Shao, J. Xie, and Y. Fang, “From handcrafted to learned and pattern recognition, 2016, pp. 770-778.
representations for human action recognition: A survey,” Image and [45] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna,
Vision Computing, vol. 55, pp. 42-52, 2016. “Rethinking the inception architecture for computer vision,” in
[24] A. B. Sargano, P. Angelov, and Z. Habib, “A comprehensive review on Proceedings of the IEEE conference on computer vision and pattern
handcrafted and learning-based action representation approaches for recognition, 2016, pp. 2818-2826.
human activity recognition,” applied sciences, vol. 7, no. 1, p. 110, [46] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. A. Alemi, “Inception-v4,
2017. inception-resnet and the impact of residual connections on learning,” in
[25] X. Cao, Z. Wang, P. Yan, and X. Li, “Transfer learning for pedestrian AAAI, 2017, vol. 4, p. 12.
detection,” Neurocomputing, vol. 100, pp. 51-57, 2013. [47] A. G. Howard et al., “Mobilenets: Efficient convolutional neural
[26] D. Wu, F. Zhu, and L. Shao, “One shot learning gesture recognition networks for mobile vision applications,” arXiv preprint
from rgbd images,” in Computer Vision and Pattern Recognition arXiv:1704.04861, 2017.
Workshops (CVPRW), 2012 IEEE Computer Society Conference on, [48] J. Chaki and R. Parekh, “Plant leaf recognition using shape based
2012, pp. 7-12: IEEE. features and neural network classifiers,” International Journal of
[27] L. Fei-Fei, “Knowledge transfer in learning to recognize visual objects Advanced Computer Science and Applications, vol. 2, no. 10, 2011.
classes,” in Proceedings of the International Conference on [49] A. R. Backes and O. M. Bruno, “Plant leaf identification using multi-
Development and Learning (ICDL), 2006, p. 11. scale fractal dimension,” in International Conference On Image Analysis
[28] H. Azizpour, A. Sharif Razavian, J. Sullivan, A. Maki, and S. Carlsson, And Processing, 2009, pp. 143-150: Springer.
“From generic to specific deep representations for visual recognition,” in [50] J. Chaki, R. Parekh, and S. Bhattacharya, “Plant leaf recognition using
Proceedings of the IEEE conference on computer vision and pattern texture and shape features with neural classifiers,” Pattern Recognition
recognition workshops, 2015, pp. 36-45. Letters, vol. 58, pp. 61-68, 2015.
[29] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman, “Return of [51] J. Hu, Z. Chen, M. Yang, R. Zhang, and Y. Cui, “A Multiscale Fusion
the devil in the details: Delving deep into convolutional nets,” arXiv Convolutional Neural Network for Plant Leaf Recognition,” IEEE
preprint arXiv:1405.3531, 2014. Signal Processing Letters, vol. 25, no. 6, pp. 853-857, 2018.
[30] M. Oquab, L. Bottou, I. Laptev, and J. Sivic, “Learning and transferring [52] E. Shelhamer, J. Long, and T. Darrell, “Fully convolutional networks for
mid-level image representations using convolutional neural networks,” semantic segmentation,” arXiv preprint arXiv:1605.06211, 2016.
in Proceedings of the IEEE conference on computer vision and pattern
recognition, 2014, pp. 1717-1724. [53] M. D. Zeiler and R. Fergus, “Visualizing and understanding
convolutional networks,” in European conference on computer vision,
[31] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature 2014, pp. 818-833: Springer.
hierarchies for accurate object detection and semantic segmentation,” in
Proceedings of the IEEE conference on computer vision and pattern [54] Y.-C. Su, T.-H. Chiu, C.-Y. Yeh, H.-F. Huang, and W. H. Hsu,
recognition, 2014, pp. 580-587. “Transfer Learning for Video Recognition with Scarce Training Data for
Deep Convolutional Neural Network,” arXiv preprint arXiv:1409.4127,
[32] H. Nam and B. Han, “Learning multi-domain convolutional neural 2014.
networks for visual tracking,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 2016, pp. 4293-4302. [55] S. G. Wu, F. S. Bao, E. Y. Xu, Y.-X. Wang, Y.-F. Chang, and Q.-L.
Xiang, “A leaf recognition algorithm for plant classification using
[33] H. Noh, S. Hong, and B. Han, “Learning deconvolution network for probabilistic neural network,” in Signal Processing and Information
semantic segmentation,” in Proceedings of the IEEE international Technology, 2007 IEEE International Symposium on, 2007, pp. 11-16:
conference on computer vision, 2015, pp. 1520-1528. IEEE.
[34] L. Shao, F. Zhu, and X. Li, “Transfer learning for visual categorization: [56] A. Aakif and M. F. Khan, “Automatic classification of plants based on
A survey,” IEEE transactions on neural networks and learning systems, their leaves,” Biosystems Engineering, vol. 139, pp. 66-75, 2015.
vol. 26, no. 5, pp. 1019-1034, 2015.
[57] A. Olsen, S. Han, B. Calvert, P. Ridd, and O. Kenny, “In situ leaf
[35] M. D. Zeiler and R. Fergus, “Visualizing and understanding classification using histograms of oriented gradients,” in Digital Image
convolutional networks,” in European conference on computer vision, Computing: Techniques and Applications (DICTA), 2015 International
2014, pp. 818-833: Springer. Conference on, 2015, pp. 1-8: IEEE.
[36] H. Azizpour, A. Sharif Razavian, J. Sullivan, A. Maki, and S. Carlsson, [58] D. Gray, S. Brennan, and H. Tao, “Evaluating appearance models for
“From generic to specific deep representations for visual recognition,” in recognition, reacquisition, and tracking,” in Proc. IEEE International
Proceedings of the IEEE conference on computer vision and pattern Workshop on Performance Evaluation for Tracking and Surveillance
recognition workshops, 2015, pp. 36-45. (PETS), 2007, vol. 3, no. 5, pp. 1-7: Citeseer.
[37] A. Sharif Razavian, H. Azizpour, J. Sullivan, and S. Carlsson, “CNN [59] J. Zhu, H. Zeng, S. Liao, Z. Lei, C. Cai, and L. X. Zheng, “Deep hybrid
features off-the-shelf: an astounding baseline for recognition,” in similarity learning for person re-identification,” IEEE Transactions on
Proceedings of the IEEE conference on computer vision and pattern Circuits and Systems for Video Technology, 2017.
recognition workshops, 2014, pp. 806-813.
[38] A. Bux, X. Wang, P. P. Angelov, and Z. Habib, “Human action
recognition using transfer learning with deep representations,” 2017.
[39] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural
information processing systems, 2012, pp. 1097-1105.
[40] C. Szegedy et al., “Going deeper with convolutions,” in Proceedings of
the IEEE conference on computer vision and pattern recognition, 2015,
pp. 1-9.
[41] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE conference on computer vision
and pattern recognition, 2016, pp. 770-778.

26

You might also like