Research On Object Detection Algorithm Of Sonar Image Based On Deep Learning

Underwater target detection,one of the prerequisites of a large number of underwater operations,distinguishes as well as locates different individuals of multiple underwater targets,guaranteeing observer’s in-depth research on underwater detection.In the meanwhile,object detection on sonar image is a crucial task of underwater target detection.Because of the high frequency,high resolution,multi-beam and the characteristics of strong real-time performance,object detection on sonar image has been widely applied in many fields such as industrial and military.Thus,object detection on sonar image is of great significance for the underwater detection.However,traditional object detection algorithm cannot meet the requirements of processing efficiency,performance or intelligence for sonar image detection.Therefore,this paper introduces deep learning technology based on convolutional neural network to satisfy the demand of high performance detection.However,the existing deep learning object detection methods fail to involve the physical characteristics as well as defects of the sonar image,which faces with two major problems.Firstly,the quality of sonar image is seriously interfered by noise,but the object detection model lacks the robustness of noise.Secondly,the available sonar image data is scarce,in the meanwhile,the physical characteristics of the acquired sonar images under various environmental conditions are quite different.Based on the two major problems above,this paper explores from the perspectives of model structure improvement and data set augmentation,carrying out research work in the following three aspects:1.Based on the analysis of the noise formation mode and characteristics of sonar images,a Noise Adversarial Network and Noise Model based on Adversarial Learning are designed.Firstly,the defects of the reference model under the sonar image target detection task are analyzed.Secondly,the noise characteristics in the sonar image generation process are investigated,which confirms sea bottom reverberation being the main external influence to produce Rayleigh noise with amplitude distribution.Then we design Noise Adversarial Network as well as Noise Block to conform the physical characteristics of sonar image.During this process,the samples of high-dimensional feature are fed into the networks to generate the noise parameters in an unsupervised way.The noise conforms to the physical characteristics of sonar image.Then the noise is induced into the original examples by multiplicative noise model of Rayleigh distribution to generate the adversarial examples.The experimental results show that the introduction of Noise Adversarial Network and Noise Block can effectively improve the accuracy and noise robustness of object detection.2.By analyzing the imaging process of sonar image,an Encoder-Decoder Based Multi-Scale Generative Adversarial Network was designed to improve the scarcity of sonar image samples.Firstly,we analyze the generation process and inversion model of sonar image.Then,according to the inversion model,the generative adversarial network which conforms to the physical characteristics of sonar image is designed.By introducing the priori knowledge of sonar image into the generative adversarial network with encode-decode process,as well as combining the features of different layers and scales in the generator,the samples conforming to the characteristics of sonar image are generated.Experiments show that the samples generated by this method are superior to the baseline method in two criteria.The accuracy of object detection can be improved by introducing the generated sample into the data set.3.Based on the analysis of the reflection model of sonar imaging,a attack method called Lambert Adversarial Sonar Attack is designed,which provides another way to implement data augmentation on object detection.First of all,based on the characteristics of Lambert surface reflection model,the original sonar image was transformed into the representation of three parameters.Since the whole process is differentiable,the original sample can be altered along the direction of gradient,which may yield misclassification result of convolution neural network.The adversarial examples are used as a supplement to the original train set to further improve the accuracy of the object detector.Finally,we summarize the characteristics and shortcomings of the above work and propose the research field and problems to be further explored.