Experimental analysis of image dehazing algorithms for pelletization process images

Abstract

The product quality of pelletization process in steel industry is usually monitored by machine vision system. However, the image quality deteriorates significantly by haze generated during pelletization. Current image dehazing algorithms mainly concentrate on natural haze in outdoor or synthetic hazy images. Whether these algorithms can be directly adopted in solving haze removal problem in industrial process images, needs to be studied. In the present work, experiments are performed to compare the performance of five state-of-the-art image dehazing algorithms, using the image dataset PELLET that consists of real hazy images captured from pelletization process in a local steel company. For a comprehensive comparative study of the image dehazing algorithms, both qualitative and quantitative evaluation criteria are adopted, including visual perceptual evaluation, no-reference image quality assessment, and taskdriven comparison. Our experimental analysis demonstrates that Boundary Constrained Context Regularization (BCCR) and Non-local (NLD) image dehazing algorithms generally achieve better quality of restored image than the other three algorithms (Dark-Channel Prior, Optimized Contrast Enhancement, and AOD deep learning network) in dealing with pelletization process images with different haze levels. The computing time needed by BCCR algorithm is only half of that by NLD (2.5 vs. 5 seconds) in processing a hazy image of size 656 × 490. Nevertheless, the performance of these algorithms needs to be improved in the future to deal with pelletization process images with dense haze, as well as to meet with the realtime requirement of pelletization process monitoring.