A motion blurring removal method based on improved DeblurGAN-v2 for coal mine belt conveyor images

In order to solve the problem that the object on the monitored image is dragged and the object contour is unclear due to the relative motion between the camera and the object on the belt of belt conveyor in coal mine. In this paper, an image motion blurring removal method based on improved DeblurGAN-v2 is proposed. Firstly, WGAN-GP is used to replace the loss function of the DeblurGAN-v2 discriminator, which suppressed the problems of unclear image contours and loss of texture details generated by the original model; Then, in order to capture deeper image details and calculate discriminator loss more accurately, a three-scale discriminator structure is designed. This allows for the better capture of the local details of generated and clear images and differences between them. The losses computed by three discriminators with different weight values are used to obtain a weighted sum. The contour of the object in the generated image is clear and the surface details are rich. But color patches unrelated to the image appear in some areas, resulting in the loss of effective information. In order to solve the color patch problem, the structural similarity loss and gradient loss are introduced into the loss function of the generator and the generator perception loss is calculated at the Conv 4-1 layer of VGG19. The generated image eliminates the local color plaque and the contour of the object in the image is clearer, the surface details are richer, and the motion blur effect is optimal. The gradient standard deviation of the generated image is 68.38% higher than that of the blurred image, which is 6 times and 17 times of the compared two methods. Finally, the foreign objects in the image before and after motion blurring removal are annotated, and these samples are trained by YOLOv5 model. The results show that the detection accuracy of foreign objects increases under different intersection ratios, with an improvement rate of 3.9% and 7.9%, respectively, and the issues of missed detection, false detection and low detection confidence of blurred images and compared methods are greatly improved. The effectiveness of the proposed image motion blurring removal method is verified.