Study On Underwater Target Recognition Technique Based On Image Content

This paper carries out with the tenth five-year national defense study-in-advance project named "underwater target acoustic detection and recognition" , which is part of the military intelligent underwater vehicle. The contents in the paper include acoustic image preprocessing, feature extraction of texture and shape, and classifier design. While solving the engineering problems, the paper makes a deep study on the recognition methods based on texture and shape features of acoustic images. The major contents are as follows in general:(1) The paper gives a survey of the internal and external current status of research and progress trends on underwater target recognition. The CBIR(Content-based image retrieval)technique, which is widely used in digital image database retrieval, is brought into the underwater acoustic image recognition.(2) To solve the disadvantages of heavy noisy disturbance and edge blurring, we thoroughly do research on relative theories and methods of morphological filtering in the pre-processing part of acoustic images. The generalized and regulated morphological filter is constructed by using multiple structuring elements. The filter possesses some important properties such as translation invariance, increasing, duality and idempotence. As the regulated operator and multiple structuring elements are used to control the morphological operation in the filter, it can not only efficiently suppress noise in images but also preserve the geometrical features of images.(3)The demerit of gray histograms is they only records the overallgray compositions of images and no any spatial information is included. To solve that, gray-spatial histograms are proposed, which incorporate spatial information with gray compositions without sacrificing the robustness and simplicity of traditional gray histograms.(4) We analyze the character of wavelet decomposition and compare the pyramid-structured wavelet decomposition with the tree-structured wavelet decomposition. In accordance to the difference between the two, the tree-structured decomposition of acoustic images has been done in the wavelet domain. Since the wavelet coefficients are shift variant, they are not suitable for direct use as texture features, which must be shift-invariant. Variance, the third moment, the fourth moment and entropy are used as texture features of images. According to the degree of dispersion, the features are normalized.(5) Comparison and analysis have been done in the selection of several typical wavelet bases which are often used in image processing. In addition, the effectiveness of texture feature sets that are made of wavelet coefficients of every level is compared. It shows that the recognition rate is not improved magnificently by using multi-level decomposition. On the contrary, the recognition rate reduces because the feature sets contain too many features and result in feature disasters.( 6 ) The recognition method of acoustic image, which takes advantages of both multi-fractal theory and wavelet analysis, is presented . In the process of feature extraction, image transformation and wavelet decomposition are combined and a feature set based on multi-fractal dimension is assembled.(7) The filled-in algorithm and Canny operator are incorporated to extract the contours of acoustic images. This method can eliminate gray burst caused by uneven echo signals. As a result, it can get integral contours and get rid of isolated points.(8) On the basis of deformable template matching, a new approach based on the deformable template is presented. Compared with the energy minimization of the Snake model, the energy function is redefined by adding a shape restriction. This improves the noise-resistance ability so that robustness and high recognizing rate are acquired. The energy minimization problem is tackled using the Immune Algorithm.(9) By introducing membership function of the features, the fuzzy classifier is designed and it reflects the uncertainty of the feature values made by random noise. This increases the robus...