Research On Target Detection Method Of Underwater Sonar Image Based On Deep Learning

Side-scan sonar has a long imaging distance and is suitable for deep-sea operations such as seabed resource exploration,seabed salvage,and underwater biometrics.It plays a very important role in underwater target detection tasks.However,due to the complex underwater environment,when using side scan sonar to carry out underwater target detection tasks,the sonar image formed has low resolution and much noise,which increases the difficulty of underwater target detection tasks.With the increasingly complex international environment and increasingly fierce competition among countries,accelerating the research on underwater target detection tasks based on side scan sonar is of great significance for strengthening China ’s maritime security and accelerating the progress of China ’s seabed resources exploration.With the advent of the era of artificial intelligence,deep learning has made great progress,and various excellent algorithm models have been gradually proposed.In view of the advantages of YOLOv5 target detection algorithm,such as high detection accuracy,strong portability and fast detection speed,it is suitable for underwater target detection tasks.Therefore,based on YOLOv5 target detection algorithm,this paper focuses on the problems existing in current underwater target detection tasks.The main research contents of this paper are as follows :(1)Sonar image preprocessing: In sonar image denoising,spatial domain filtering and transform domain filtering are used to complete,and peak signal-to-noise ratio,mean square error,and structural similarity are introduced as image quality evaluation indicators.Among them,in view of the problem that the traditional median filter makes the sonar image after denoising too smooth,the median filter algorithm is improved,and a sonar image denoising method based on a new median filter is proposed.The experimental results show that: the new method proposed in this paper The median filter has the best denoising effect.In sonar image enhancement,the histogram equalization,histogram regulation,and single-scale Retinex algorithm are used to complete,and the human visual system,contrast,and information entropy are introduced as the quality evaluation indicators of the enhanced sonar image.From the experimental results,it can be known that: Scale Retinex algorithm has the best image enhancement effect.(2)Improve the YOLOv5 target detection algorithm: use the K-means++ clustering algorithm to reselect the anchor box clustering method to avoid the occurrence of weak clustering problems;use the mixup algorithm to expand the sonar image data set online to enhance the richness of the data set degree;use Focal-EIOU Loss as the bounding box regression loss function to speed up model convergence;use Soft-NMS algorithm to screen the final generated bounding box to solve the problem of false detection;introduce ECA attention mechanism into the model to avoid model degradation Dimension;use the Bi FPN structure to achieve efficient multi-scale fusion of information extracted from features.In this paper,the improved YOLOv5 target detection algorithm is trained and tested on the self-made data set.The experimental results are as follows : the accuracy is 97.6 %,the regression rate is 97.7 %,m AP @.5 is 98.1 %,m AP @.5 :.95 is 73.4 %.Compared with the original YOLOv5 algorithm,the accuracy is improved by 6.3 percentage points,the regression rate is improved by 6.2 percentage points,m AP @.5 is improved by 6percentage points,m AP @.5 :.95 is improved by 4.4 percentage points.The improved algorithm has obvious performance improvement,which proves that the proposed method has certain feasibility and effectiveness in underwater target detection tasks.