Study On Underwater Target Recognition Technique

With the development of the acoustic imaging technology, sonar image has been widely used in the ocean exploitation. The target identification based on sonar image has become an important subject in the digital image processing domain. This thesis makes a deep study on recognition methods based on texture and shape features of acoustic images. The main contents and contributions of this thesis are showed as follows:(1) According to the characters of torpedo, naval mine and submarine, this thesis extracts the shape features as NMI feature, invariant moments and relative moments. As these features have the rotation invariance, scale invariance and transfer invariance, so they can be used to recognize the underwater targets.(2) The conventional gray histogram method is lack of the ability to describe the spatial information of the image. Considering the characteristics of sonar image, this paper presents a new gray histogram, using the roundness of the prominent region as the image feature. The amalgamation of gray, spatial, shape information improves the efficiency of sonar image recognition.(3) The wavelet transform is employed to extract the texture features of sonar image. This thesis selects several typical wavelet bases which are often used in image processing to compare and analyze. Then discusses the relationship between wavelet bases and the recognition rate. Since the wavelet coefficients are shift variant, they are not suitable for directly used as the features. So mean, variance and entropy are used as texture features.(4) The traditional wavelet transform needs complex computation and large memory size. Considering the characters of sonar image, lifting scheme is used to construct the Haar wavelet and a new method of wavelet decomposition is presented. The experimental results show that the new method has higher classification rate and is more efficient than traditional wavelet.(5) The fractal dimension lacks of the ability to describe the spatial information of the image. Considering the characteristic of sonar image, this paper uses lifting scheme to construct the Haar wavelet and associated lifting scheme with fractal dimension. The amalgamation of multi-scales character of wavelet transform and fractal dimension increases the recognition rate. BP neural network is used to recognize the sonar images of different SNR. The results show that the new arithmetic is effective.(6) Levenberg-Marquardt (L-M) algorithm which has second-order convergence effects is utilized to optimize standard BP algorithm and it is applied to the sonar image recognition systems. The experimental results indicate that the LMBP neural network has higher classification rate, more rapidly converge velocity and better anti-noise performance by comparing with other improved BP algorithm and the RBF neural network.