| abstract |
The invention provides an underwater image denoising method based on a generative confrontation network. First, the underwater image is input into a generative network composed of several residual blocks to obtain a feature map; then the feature map obtained by output is combined with The noise-free label images of Shimizu are mapped through the VGG‑19 network to obtain a deep feature space, and the perceptual cost of the feature map and the noise-free label image of Shimizu in the depth feature space is calculated. While calculating the perceptual cost, the network output is generated. The feature map is input to the adversarial network, and after the final training is completed, the noisy underwater image is input to the generation network, and the output is the processed noise-free image. By introducing the confrontation mechanism, the invention has obvious denoising effect, especially the method can effectively retain or even enhance the edge texture information in the image, and has better visual effect and imaging quality. |