Target Detection And Data Augmentation Method Based On Deep Learning For Sonar Image

Sonar is a general term for devices that use sound waves as an information carrier and is an important tool for locating,detecting,tracking,and identifying targets in the marine environment.In recent years,with the development of underwater acoustic imaging detection technology,using sonar images to detect underwater targets has become a key means.Early sonar image target detection technology relied on manual experience and expert supervision,which was quite time-consuming and resource-intensive.Using artificial intelligence to track and detect underwater acoustic signals and using machines to extract features from sonar images and perform target detection can greatly improve the technical level of the sonar image target detection field.This article uses machine learning and deep learning methods to detect targets in sonar images based on their characteristics.In response to the problems of sample scarcity and uneven distribution in sonar images,a generative adversarial network structure is proposed to achieve the generation of sonar images.The specific research content is as follows:First,using a combination of artificial feature extraction and machine learning for sonar image object detection.Perform denoising,grayscale enhancement,segmentation,HOG feature extraction,and SVM object detection combined with sliding windows on sonar images.Analyze the advantages and disadvantages of this method based on experimental results.Then,use YOLOv3 and transfer learning algorithm for sonar image target detection.Analyze the working principle and network structure of YOLOv3,introduce transfer learning method to increase the generalization ability of the model and improve the target detection accuracy.The experimental results show that the average detection accuracy using YOLOv3 is0.799.Finally,to address the issues of sample scarcity and uneven distribution in sonar images,an improved DCGAN network structure is proposed for data enhancement of sonar images.This network can directly convert input noise into generated sonar images.Comparing the network structures of the original GAN,DCGAN,and improved DCGAN,and conducting experiments,it is demonstrated that the improved DCGAN can generate higher quality and resolution sonar images.Using improved DCGAN generated sonar images as an expanded dataset to train the YOLOv3 detection model can improve the average detection accuracy of the model to 0.841.