Research On Techniques Of Sonar Image De-noising And Segmentation

In acoustic detection, sonar image is the data visualization result of the echo from targets, while sonar image de-noising and segmentation play a key role in later target recognition. Based on acoustic imaging theory, sonar image has its own characteristics, so there are certain particular conditions or demands for de-noising and segmentation. In this sense, it is worth researching suitable processing methods for sonar image, though there have been many reliable optical image de-noising and segmentation algorithms. In this paper, some significative methods for sonar image de-noising and segmentation were discussed.For the sonar image de-noising, more popular wavelet analysis was selected as the theoretical basis. The major researches of this paper were two relatively new wavelet frameworks. One of which took direction analysis into account, called directional wavelet transform, and the other took morphological analysis into account, called morphological wavelet transform. Among these wavelet transforms, some had been proposed years ago, or some were very mature, but rarely applied and contrasted in sonar image processing. So in this paper, through a large number of simulation experiments, the performances of these wavelet transforms in sonar image denoising were discussed respectively. And under the morphological wavelet signal decomposition framework, specific expressions of morphological wavelet were proposed on the basis of order-statistics filters and mean filter, then using the subdivision and lifting steps to improve the morphological wavelet performance. After summing up the sonar image de-noising algorithm with morphological wavelets, de-noising experiments were made on some classical sonar images. Compared with other wavelet de-noising algorithms, our method was better, faster and more stable. Then, on the basis of mean morphological wavelet proposed, we tried the coefficient modeling. We analyzed the distribution of coefficients in morphological mean wavelet domain to construct a Hidden Markov Tree model, and proposed the corresponding de-noising algorithm. Finally, algorithm simplification was also studied and some valuable conclusions were obtained. Applying morphological mean wavelet HMT model de-noising method in sonar image processing, simulation experiment results showed the validity of our model compared with traditional wavelet HMT model de-noising and popular Contourlet HMT model de-noising.For the sonar image segmentation, this paper was also divided into two parts to start the research. Firstly discuss the application of level set theory in sonar image segmentation, and some region-based active contour segmentation models were introduced, including PC model, PS model, as well as LBF energy model. On the basis of Li's LBF energy model, this paper made some improvements according to the characteristics of sonar image. We gave a method to define the initial contour and proposed to use Laplacian for determining the level set evolution direction. Simulation results showed that these improvements had great significance for the sonar image segmentation. By which the segmentation was not only non-sensitive to noise, but also accurate when sonar image had poor edge, complex contents or even shadows. Secondly, this paper introduced graph theory into sonar image segmentation, and paid attention to normalized cut algorithm, graph cuts without eigenvectors, standard spectral clustering and self-tuning spectral clustering respectively. And on the basis of self-tuning spectral clustering, this paper proposed a fast and unsupervised method for sonar image segmentation. Through some preprocessing steps and an improved affinity function, the spectral clustering segmentation method proposed in this paper was suitable for sonar image processing, and achieved unsupervision basically.