Sonar Image Target Detection Based On Multi-scale And Multi-feature Fusion

With the development of ocean engineering,seabed mapping has become the main content of ocean research.Seabed sonar detection technology is an important basic technology for marine exploration.The analysis and processing of sonar image have been widely concerned by the academic circles at home and abroad.For data processing method of target detection in sonar image.On the basis of previous research work,based on the pixel classification method of Sonar image,this thesis explores and realizes the target detection of Sonar image.This thesis mainly studies the following aspects:(1)The denoising method of sonar image is studied.Aiming at the problem of speckle noise in sonar image which is difficult to remove,the 3D matching block filtering method based on wavelet decomposition is analyzed and implemented in this thesis.In this method,the noise of the sonar image is firstly modeled by Gaussian noise,and the high-frequency components of the image with noise are obtained by wavelet decomposition.Then,the high-frequency components are removed by 3D matched block filtering method.Finally,the image after denoising is reconstructed by wavelet.The experimental results show that this method can remove the noise and retain the edge details of the image.(2)The low-level feature extraction method of sonar image is studied.In order to obtain the fractal features describing the texture information of sonar image,According to the dramatic change of the fractal features of the image target when the scale changes,a fast carpet image fractal feature extraction method based on Gaussian pyramid is proposed in this thesis.In order to obtain the contrast feature describing the sonar image gray level information,in view of the problem of large contrast errors in the image edge region obtained by traditional methods,the contrast features of sonar images are obtained by calculating the image gray similarity.In order to get the edge density features describing the edge information of sonar image,according to the feature that the target region of the image has abundant edge features compared with the background region,the edge density feature extraction method of sonar image based on voting matrix is proposed.The experimental results show that the feature extraction method in this paper is accurate and improves the classification accuracy.(3)Based on conditional random field,a full convolution network sonar image feature extraction method is analyzed and implemented in this thesis.The full convolution network model is used to obtain the sonar image features and input the features into the conditional random field model to obtain the final sonar image features.The experimental results show that this method can obtain fine image features.(4)The target detection method for sonar image is studied.In this thesis,SVM classifier is used to classify pixels to realize sonar image target detection,and genetic algorithm is used to optimize the classifier parameters.The experimental results show that this method can remove the influence of background and detect the target accurately and effectively.