Improved U-Net for Industrial Image Denoising
Peng Zhang, Wenhua Jiao, Yinqiang Zhang, Lijuan Li, Dengfeng Zhang · 2023
Many cameras are now used in industrial settings to monitor production processes. The cameras inevitably bring in noise while transmitting the images, which hinders the understanding of the images for subsequent vision tasks, such as worker detection. Researchers usually use methods based on deep learning to denoise industrial images, which tend to corrupt the texture and edge information of the images. To solve this problem, this paper proposes a U-Net algorithm based on feature fusion and adaptive fusion interpolation. First, a dense convolution module is added to the feature extraction process to improve the utilization of different levels of features, and then the images of different sizes are decoded using up-sampling based on adaptive fusion interpolation to recover clean images. The experimental results show that the proposed algorithm in this paper is more effective compared with existing algorithms and can better preserve texture and edge information while denoising.