Research On Sonar Image Recognition Based On Artificial And Non-Artificial Feature

This work concentrates on two types of feature extraction and classification methods for sonar images.For a specified target,a suitable texture feature vector for extraction is artificially designed and then putted into a classifier to distinguish objects;For a multi-kind task,a convolution neural network is used.The first part of this work has focused on image preprocessing,including image denoising and image segmentation.For the denoising part,neighborhood averaging method,wiener filtering,wavelet filtering and BM3 D are juxtaposed,and BM3 D is proved to be the suitable algorithm for sonar image denoising.For the segmentation part,four algorithms are compared for specified targets,which are iteration method,Otsu method,maximum entropy method and fast marching method.By measuring the segmentation result,iteration method proves to be suitable.Based on the work above,feature extraction and classification have been researched.On the bi-classification situation,a feature vector made up by image texture features,computed by GLCM,is constructed first,then the vector is passed to a detractor,which is SVM,to classify.Simulation result proves the efficiency of this model.For the multi-kind problem,a convolution neural network architecture of two depth is constructed,which uses convolution kernel to compute feature maps and BP algorithm to adjust the link parameters instead of depending on human designer to decide which features to be extracted.The simulation result proves the effectiveness of this architecture.