Introduction: Object Detection is a computer vision task in which the goal is to detect and locate objects of interest in an image or video. The task involves identifying the position and boundaries of objects in an image, and classifying the objects into different categories. It forms a crucial part of vision recognition, alongside image classification and retrieval. Finding objects in images taken in bad weather is tough for existing object detection systems. CNN-based methods have become prevailed in object detection. They not only have achieved promising performance on benchmark datasets but also have been deployed in the real-world applications such as autonomous driving. Due to the domain shift in input images, general object detection models trained by high quality images often fail to achieve satisfactory results under adverse weather conditions (e.g., foggy and dark light). In the context of adverse weather conditions, image degradation often includes effects like fog, rain, snow, or low-light conditions that can obscure details, reduce visibility, and distort the appearance of objects in the image. This degradation makes it challenging for existing computer vision systems to accurately detect and recognize objects within these images.
ARTIFICIAL NEURAL NETWORKS 20/12/2023 3
Problem Statement: Deep learning-based object detection methods have achieved promising results on the conventional datasets, but it is still challenging to locate objects from the low-quality images captured in adverse weather conditions. The existing methods either have difficulties in balancing the tasks of image enhancement and object detection, or often ignore the latent information beneficial for detection. Proposed Solution: The proposed IA-YOLO method aims to address the challenge of poor visibility in images taken in adverse weather conditions, which hampers object detection. The approach involves an image-adaptive detection framework designed to remove weather-specific information and reveal more hidden details. Image-Adaptive YOLO consist of three components;
Predictor (CNN-PP): (DIP) Module: This component predicts The input image, filtered by the DIP The processed image from parameters for a Differentiable module using the predicted the DIP module is then used Image Processing module (DIP) parameters, serves as the input for as input for object detection based on the input image. It takes the YOLOv3 detector. The DIP the resized image (256×256) as aims to remove weather-specific using the YOLOv3 network. input and estimates the parameters information, enhancing the image needed for image processing. for better object detection.
ARTIFICIAL NEURAL NETWORKS 20/11/20XX 5
ARTIFICIAL NEURAL NETWORKS 20/12/2023 6 The DIP module comprises six differentiable filters with adjustable hyperparameters. These six filters are divided into 3 categories:
Defog Filter (ω) specifically designed to handle foggy scenes.
Standard Filter include White Balance (WB), Gamma, Contrast, and Tone, which operate pixel-wise and can be expressed as pixel-wise filters. Sharpen Filter (λ) enhances image sharpness, potentially improving object details and boundaries.
ARTIFICIAL NEURAL NETWORKS 20/12/2023 7
STANDARD FILTERS: The pixel-wise filters map an input pixel value Pi = (ri,gi,bi) into an output pixel value Po = (ro,go,bo), in which (r,g,b) represent the values of the three color channels red, green and blue, respectively.. Equations for Contrast and tone filters:
ARTIFICIAL NEURAL NETWORKS 20/11/20XX 9
Dataset:
Experiments conduct on three datasets:
⮚ VOC_norm_trainval ⮚ VOC_norm_test ⮚ RTTS(comprehensive real-world dataset available in foggy conditions)
ARTIFICIAL NEURAL NETWORKS 20/12/2023 10
Experimental Results: To evaluate the effectiveness of method in the scenarios of fog and low-light. The filter combination is [Defog, White Balance(WB), Gamma, Contrast, Tone, Shapen], while the Defog filter is only used in foggy conditions.
Experiments on Foggy Images:
Table compares the mean average precision (MAP) results between IA-YOLO and other Competing algorithms in both the normal and hazy conditions.
ARTIFICIAL NEURAL NETWORKS 20/12/2023 11
Cont… Below fig shows several visual examples of our IA-YOLO method and the baseline YOLOv3 II. Though for some in stances, our adaptive DIP module produces some unwilling noises for visual perception, it greatly boosts the local image gradients based on image semantics, and leads to better detection performance.
ARTIFICIAL NEURAL NETWORKS 20/12/2023 12
Cont… Experiments on Low-light Images: We compare our presented IAYOLO method with baseline YOLOv3, Enhance + Detect, DAYOLO, and DSNet on the three testing datasets. For Enhanc + Detect, we employ the recent image enhancement method Zero-DCE to preprocess the low-light images and use YOLOv3 trained on VOC_norm for detection. The remaining experimental settings are kept the same as those on foggy images.
ARTIFICIAL NEURAL NETWORKS 20/12/2023 13
Cont… Fig. 4 shows the qualitative comparisons between IA YOLO and the baseline YOLOv3 II. It can be observed that proposed DIP module is able to adaptively increase the contrast of input image and reveal the image details, which are essential to object detection
ARTIFICIAL NEURAL NETWORKS 20/12/2023 14
Conclusion: IA-YOLO approach significantly advances object detection in adverse weather conditions. By adaptively enhancing each input image through a differentiable image processing module, we effectively removed weather-specific information, restoring crucial content for optimal YOLO detector performance. Training the entire framework end-to-end allowed the parameter prediction network to weakly supervise and learn the ideal DIP module via the detection loss. Proposed method showcased remarkable adaptability in handling both normal and adverse weather conditions. Experimental results demonstrated its superiority over prior methods, particularly in foggy and low-light scenarios. The IA-YOLO approach holds immense promise for substantially improving object detection reliability in challenging weather conditions, benefiting various real- world applications.
ARTIFICIAL NEURAL NETWORKS 20/12/2023 15
THANK YOU ARTIFICIAL NEURAL NETWORKS 20/12/2023 16