​Port Conveyor Belt Defect Detection Enhanced by CNN Dehazing: An Industrial Validation Using Faster R-CNN​

Haze weather has a great negative impact on image processing, while traditional haze image restoration methods are based on certain assumptions or a priori conditions, which limit the wide application of the algorithm. To address these challenges, we introduce a CNN-enhanced framework for single-image haze removal. The model incorporates dark-channel–guided feature cues to refine the CNN architecture, subsequently enabling accurate estimation of scene transmission. At the same time, the atmospheric light value A of the haze image is estimated by the optimized atmospheric light value estimation method, and finally the restoration model of the haze image is constructed, and the restoration of the haze image can be realized by the model. In order to verify the image restoration effect of the above method, this paper uses Faster-RCNN to perform target detection on the dehazing image for objectivity evaluation. Finally, the experiment is carried out by using natural haze images and artificially synthesized haze images, and with other kinds of dehazing. Comparative evaluations demonstrate that our approach consistently outperforms existing dehazing techniques.