Study on Practical Utility of Image Dehazing Algorithms Based on Deep Learning in Computer Vision Scene Understanding

Abstract

Deep learning-based methods have achieved remarkable performance for single image dehazing. However, previous studies are mostly focused on the dehazing of human-perspective images without considering the practical application in the field of computer vision. Whether the image dehazing of deep learning-based methods(IDDL) can be applied in the field of computer vision will determine the research direction of computer vision dehazing scenes. This work presents a comprehensive study that the feasibility and reliability of applying IDDL from human vision to practical computer vision scenarios. An approach for practical utility and portability evaluation of the IDDL based on human vision, including the comparison regarding the detection accuracy, model running time, and model scale of the detection method, is proposed. Our research focuses on re-evaluating the applicability of image dehazing algorithms based on a computer vision perspective, including evaluation method and evaluation index. To better verify the detection effect of the dehazing algorithm, we use the synthetic dataset method as a control experiment and make a comparison with 6 state-of-the-art dehazing methods on a real-world hazy scene. Additionally, investigate the proportion of synthetic hazy images in the training dataset used, including object detection and semantic segmentation algorithms. Extensive experiments demonstrate that the poor practical utility and portability of IDDL in computer vision are verifiable, and the IDDL solution should be extensively paid attention to concerning the computer vision domain rather than leveraging the IDDL based on human vision. Finally, We provide some methods to solve the hazy scene.