Research On Target Detection Technique Of Underwater Sonar Image

Autonomous target detection and recognition based on underwater sonar shows its significance and value in both military and civilian fields with the continuous development of imaging sonar systems and the emergence of underwater intelligent robots. In this paper, a large number of experiments were carried out on the underwater target detection and recognition system with using different target detection and segmentation algorithms. Based on these, we studied the image data pre-processing image segmentation and sonar image feature extraction. And we designed corresponding real-time target detection systems. The main work is as follows.Firstly, we learned that sonar image noise pollution is serious in the most sonar images through the analysis in the pre-processing of the sonar images. On this basis, both the spatial domain and frequency domain denoising methods of underwater sonar images are analyzed and experimented. And a Gaussian pyramid method based on multi-resolution sonar image filtering is applied to the pre-noising, and it achieves relatively good results.Secondly, taking the sonar image characteristics into account, we optimize an unsupervised automatic sonar image segmentation method based on MRF (Markov Random Field). After analyzed a large number of sonar images, we find that the reverberation area of sonar images basically obey the rule of Gamma distribution. In this case, we use a fast and effective Gaussian Pyramid model for the sonar image preprocessing to make the underwater reverberation of these images obey Gaussian distribution. On this basis, we design a kind of model to classify the sonar images automatically. And we use a local energy maximum method to estimate the initialization parameters of MRF model. Then a fully automated sonar image segmentation model was formed by these two steps. Also, we design experiments to test this model. Compared with other typical sonar image segmentation algorithm, this method is effective and fast.Thirdly, after having analyzed the texture features, shape features and feature extraction techniques of sonar image, a sonar image feature extraction method focusing on a new class-based Haar features are designed. The experimental analysis and results show that the effect of image feature extraction is improved obviously by this method.Finally, based on the above researches, corresponding real-time target detection systems were designed for the Klein 5000 sidescan sonar and Blue View forward looking sonar. And a research on Haar features class-based object detection system is also carried out. For the problem that the same detection algorithms cannot apply to targets in all frames on Blue View forward looking sonar target detection system, which causes missing target in some jframes on particular situations. This paper uses a Kalman prediction associated with target detection technology. i.e., we combine the previously target detection results and predictions based on Kalman comprehensive goals in order to get the final test results. Experiments show that the method works better.