A Method of Removing Oil Droplets from Bearing Image Based on a Two-stage Neural Network

In industrial visual defect detection, the phenomenon of light emission caused by surface oil will seriously reduce the imaging effect, leading to a high false detection rate of the defect detection algorithm. This paper proposed an algorithm based on a two-stage network to remove oil droplets from the bearing image. The proposed algorithm can effectively remove oil droplets from the image and significantly improve surface defect detection accuracy. We introduced the channel attention mechanism and improved Attentive-GAN to improve detail recovery ability in the raindrop area. We use the Convolutional Block Attention Module to improve the generator network of AE-WGA, solving the texture detail ambiguity caused by the raindrop removal algorithm and restoring high-resolution surface details. Experimental results show that the proposed algorithm can achieve a PSNR of 31.24437 and SSIM of 0.93926 and reduce the error detection rate of the pit defect detection model by 93%.