Remove and recover: two stage convolutional autoencoder based sonar image enhancement algorithm

High-quality forward-looking sonar images are the basic guarantee for underwater object detection and classification of autonomous underwater vehicle (AUV). However, the sonar image has been suffering from two main problem with complex and changeable underwater environment: high speckle noises and the lack of high frequency information. In this paper, a two-stage sonar image enhancement algorithm based on convolutional autoencoder is proposed to solve the above two problems to allow low-frequency sonar images to obtain resolutions approximate to high-frequency sonar images. For the high speckle noise, we proposed a convolutional denoising autoencoder based speckle reduction method for low-frequency sonar image to avoid noise enhanced as the image enhancement process. Skip connections and newly designed loss function is incorporated to better suppress noise with varying degrees. To solve the problem of insufficient high frequency information, a convolutional sparse autoencoder is further introduced to achieve image super resolution enhancement. In order to verify the effectiveness of the enhancement network proposed in this paper, we conducted extensive experimental analysis have been conducted from two aspects: image enhancement effect and underwater target detection effect. Specifically, underwater sonar images have been collected was conducted based on our self-owned AUV platform equipped with a dual frequency forward looking sonar in a water tank for network training. And three datasets are constructed for speckle reduction, high-frequency restoration, and underwater target detection. Through extensive experimental verification, our proposed enhancement method achieves high performance improvement on speckle reduction and high-frequency information restoration than the state-of-the-art image speckle reduction and image enhancement algorithms with a better PSNR, SSIM, and EPI. Finally, the YOLO V5 network model is used for underwater target detection and the experimental results show that combining image enhancement networks can effectively alleviate the problems of false and missed detections in the original network, which greatly improving the detection accuracy of the algorithm with a high detection speed. All the experiments have shown that the method proposed in this paper can effectively improve the underwater perception ability of sonar equipment.