An Image Enhancement Method for Side-Scan Sonar Image Based on Multi-Stage Repairing Image Fusion
The noise interference of side-scan sonar images is stronger than that of optical images, and the gray level is uneven. To solve this problem, we propose a side-scan sonar image enhancement method based on multi-stage repairing image fusion. Firstly, to remove the environmental noise in the sonar image, we perform adaptive Gaussian smoothing on the original image and the weighted average grayscale image. Then, the smoothed images are all processed through multi-stage image repair. The multi-stage repair network consists of three stages. The first two stages consist of a novel encoder–decoder architecture to extract multi-scale contextual features, and the third stage uses a network based on the resolution of the original inputs to generate spatially accurate outputs. Each phase is not a simple stack. Between each phase, the supervised attention module (SAM) improves the repair results of the previous phase and passes them to the next phase. At the same time, the multi-scale cross-stage feature fusion mechanism (MCFF) is used to complete the information lost in the repair process. Finally, to correct the gray level, we propose a pixel-weighted fusion method based on the unsupervised color correction method (UCM), which performs weighted pixel fusion between the RGB image processed by the UCM algorithm and the gray-level image. Compared with the algorithm with the SOTA methods on datasets, our method shows that the peak signal-to-noise ratio (PSNR) is increased by 26.58%, the structural similarity (SSIM) is increased by 0.68%, and the mean square error (MSE) is decreased by 65.02% on average. In addition, the processed image is balanced in terms of image chromaticity, image contrast, and saturation, and the grayscale is balanced to match human visual perception.
National Science Foundation of China (NSFC) under Grants 62203133, the National Key R&D Program of China (2021YFB3901300) and the State Key Laboratory of Robotics and System (SKLRS-2023-KF-17).
Author affiliationSchool of Computing and Mathematical Sciences, University of Leicester
- VoR (Version of Record)