Research On Image Segmentation Based On Parzen Window And Q-learning

Image segmentation is an essential step from image processing to image analysis that can divides the image into numbers of region of significance without overlap according to its characteristics. Image segmentation is regarded as the bottleneck in the procedure of image processing, the quality of which has direct effects on the follow-up work. Therefore, the research and development of image segmentation have very significant importance.Image segmentation is one of the most important and critical image analysis techniques for target extraction and measurement. The result of image segmentation is the foundation for feature extraction and recognition, which has been the hot focus in the image technology researches. Issues of image segmentation has been studied as well as implemented in this thesis.A comprehensive method has been brought out that the using of the threshold segmentation method of Parzen window estimate based on average gray neighborhood information is considered into average neighborhood pixel gray-scale information as well, so as to obtain a better segmentation result when the SNR is relatively low or the image background is more complicated. However, the proposed algorithm should traverse the search space of two-dimensional gray-scale image, which lead to a high algorithm complexity and long computing time. To solve this problem, the transition zone is used to shorten the computing time. At the same time, in order to further reduce the computing time of proposed algorithm in this thesis. We also used the particle swarm optimization (PSO) to accelerate the search speed. The experimental results show that the proposed algorithm will be able to obtain a better segmentation result in images of poor quality, and can meet the general requirements of real-time in the general environment.Taking into account the complementarity of different theory-based threshold segmentation algorithms in image segmentation, this thesis purposed an algorithm based on Q-learning convergence threshold segmentation algorithm that integrates threshold segmentation algorithms of different theories, so as to improve the of general adaptability for the algorithm. The purposed algorithm considered three types of evaluation index for image segmentation quality (such as inter-regional contrast, intra-regional homogeneity, shape measurement) to construct the accumulation function returns, and adjust the weights of different segmentation methods in order to get the maximal return, so that the proposed algorithm can achieve better segmentation results and general adaptability. Meanwhile, this algorithm also uses the particle swarm optimization (PSO) to accelerate the search speed. The experimental results show that the proposed threshold segmentation algorithm based on Q learning not only possesses of better performance than single threshold segmentation algorithm, but also have better general adaptability.Finally, all the research work has been summarized in the thesis, and brings forward the direction of further research.