Research On Underwater Target Detction Method Based On Deep Learning

The combination of deep learning and underwater target detection technology has brought unprecedented new opportunities to underwater target detection tasks,which will play an important role in both the military and civilian fields.This article aims at the problems of poor detection speed and insufficient recognition accuracy of existing target detection algorithms for underwater target detection.According to the improved deep learning target detection algorithm to improve the speed and accuracy of underwater target detection.So that the algorithm can achieve ideal results in practical applications.The main work of this paper is as follows:(1)This paper introduces the basic concepts commonly used in deep learning,as well as some classic convolutional neural network models and deep learning target detection models.Analyzes the advantages of deep learning target detection algorithms in the field of underwater target detection.Through the comparison of speed and accuracy,the YOLOv4 model is selected as the research basis for developing an underwater target detection model,which is suitable for the needs of underwater target detection.(2)Underwater targets are affected by the complex underwater environment,which can cause problems such as blurred images,color inconsistencies,and noise interference in the collected images.This can lead to existing deep learning target detection networks not being able to fully extract effective information from underwater targets.For this reason,this article improves and optimizes the YOLOv4 model,by adding a self-designed parallel space and channel attention module PDAM,simultaneously extracting spatial and channel information from feature maps,to enhancing the network’s ability to extract features from underwater targets.And adaptively combining the local features of underwater targets with their global dependencies,capturing the intercorrelations between features,thus effectively improving the accuracy of the algorithm for detecting underwater targets.Secondly,three network model are constructed,YOLOv4-PDAM-1,YOLOv4-PDAM-2 and YOLOv4-PDAM-3,were designed by embedding PDAM module in different positions of the YOLOv4 network.Through ablation experiments,the impact of each improvement model on underwater target detection performance was analyzed,and compared with the detection performance of the original YOLOv4 model,thereby proving the effectiveness of the PDAM module and the improved network.(3)In order to the slow real-time performance of YOLOv4 model in underwater target detection,this paper adjusts the structure of the YOLOv4 model,replaces the CSPDarknet53 network in the original YOLOv4 with a lightweight network Mobile Netv2,and introduces deep convolution and point-wise convolution to reduce computational redundancy,thereby obtaining the M2-YOLOV4 network model,which can greatly improve the speed of underwater target detection,making the model meet the requirements of real-time performance.At the same time,Coordinate Attention Atrous Spatial Pyramid Pooling CAASPP module is also proposed,which consists of Coordinate Attention CA module an Atrous Spatial Pyramid Pooling module ASPP.Introducing the CAASPP module into the M2-YOLOv4 network to replace the original SPP module can alleviate the accuracy degradation caused by the small number of layers of the lightweight network Mobile Netv2,making the final formed M2-YOLOv4+CAASPP model take into account both detection speed and detection accuracy,and has improved performance of underwater target detection.