Research On Underwater Optical Image Target Detection Algorithm Based On Deep Learning

With the development of technology,underwater robots equipped with intelligent vision systems have become an essential tool for effective exploration and exploitation of marine resources.Underwater image enhancement technology is particularly important in order to better perceive the underwater environment for efficient operation of underwater robots.In addition to this underwater robots need high accuracy and speed of underwater target detection techniques.Therefore,this paper addresses underwater image enhancement techniques and underwater target detection techniques,with the following main work:For the problems of color recession and low contrast of underwater images,based on Multi-Scale Retinex with Color Restore(MSRCR)algorithm and on Global Histogram Stretching(RGHS)algorithm,this paper proposes MSRCR-RGHS underwater based on Laplace pyramid fusion image enhancement algorithm based on Laplace pyramid fusion.First,the MSRCR algorithm with bilateral filtering and improved color recovery factor is used to solve the problem of color recession of underwater images;then,the RGHS algorithm with gamma correction is used to solve the problem of low contrast of underwater images;finally,the images processed by the improved MSRCR and RGHS algorithms are multi-scale fused based on the Laplace pyramid to obtain the final enhanced images.To address the problems of difficult underwater small target detection and lightweight underwater detection model,an underwater target detection algorithm based on YOLOv7 and depth separable convolution(ND-GS-YOLOv7)is proposed in this paper.The algorithm firstly replaces the ELAN structure in the backbone network with the New DSCLayer structure to increase the perceptual field to improve the detection accuracy of overlapping targets;secondly uses the Sim AM attention mechanism to improve the detection accuracy of small targets;then replaces the ELAN structure in the detection head with the GSCLayer structure to improve the real-time target detection;on top of that,uses the SIo U loss function to solve the problem that underwater targets have directionality;finally,the Grad-CAM visual heat map framework is introduced to validate the improved part.In this paper,we use the 2020 underwater optical target detection dataset,and firstly,we compare whether the dataset is processed with image enhancement algorithm or not,and the results show that the accuracy of target detection is higher after image enhancement;then we conduct ablation experiments on ND-GS-YOLOv7 target detection algorithm to verify the necessity of algorithm improvement;finally,we use mean average accuracy m AP and detection speed FPS as performance evaluation indexes to compare with Finally,the performance evaluation indexes of mean average precision m AP and detection speed FPS are compared with other algorithms,and the experiments show that the algorithm in this paper achieves a balance between detection accuracy and detection speed,and improves the detection accuracy while basically not reducing the detection speed,which has better application prospects.