LightNetDehaze: Lightweight and Unsupervised Single-Image Dense Haze Removal

<div>Image dehazing techniques can play a vital role in object detection, surveillance, and accident prevention, especially in scenarios where visibility is compromised because of light scattering by atmospheric particles. To obtain a high-quality image or as an initial step in processing, it’s crucial to restore the scene’s information from a single image, given that this is an ill-posed inverse problem. The present approach utilized an unsupervised learning approach to predict the transmission map from a hazy image and used YOLOv8n to detect the car from a clear recovered image. The dehazing model utilized a lightweight parallel channel architecture to extract features from the input image and estimate the transmission map. The clear image is recovered using an atmospheric scattering model and given to the YOLOv8n for car detection. By incorporating dark channel prior loss during training, the model eliminates the need for a paired dataset. The proposed dehazing model with fewer parameters speeds up the dehazing process, which can detect the objects in less response time. The network follows unsupervised learning, which eliminates the need of ground truth image or transmission map of a clear image. The proposed method tried to solve the issue of high computational complexity and long latency when used as a preprocessing stage in computer vision applications. The proposed network ranks first in terms of parameters and FLOPs, which are lower by scale 10<sup>2</sup> and 10<sup>3</sup>, respectively, compared to the method ranked second. The results highlight the effectiveness of the proposed method compared to other methods and ranked first in number of car detections using YOLOv8n. The inference time to dehaze the image is comparable to the method ranked first and 66% lower than the third rank.</div>