Image Processing Basing On Fuzzy Theory And Hidden Markov Model

In recent years, image processing is one of the fastest growing areas, its application has penetrated all aspects of social life. During the process of the image processing, image segmentation is the first task in image analysis, pattern recognition, computer vision and other areas. Fuzzy clustering can describe the intermediary of the sample classes ,and can reflect the real world more objectively, it has become the mainstream of cluster analysis. Among many fuzzy clustering methods, the most widely used is the FCM algorithm, but there are still many problems to be solved. These shortcomings result in a bottleneck of image segmentation, which restricts further development in this area. In the field of image recognition, evolution from the pixel-based to stroke-based image analysis is requirement for further development. In the late 20th century, Hidden Markov Model was applicated in speech recognition, Gesture recognition, handwritten document understanding and many other areas. But the classical algorithm for training HMM is the baum-welch algorithm, it is actually a maximum likelihood parameter estimation method. Due to the single model evolutes independently, converge gradually, it can only get local optimization,but to find the optimal HMM is very difficult.Based on the conclusion of existing research fruit, this paper proposes some modified methods, specific tasks are as follows:1. A new method for remote sensing image segmentation based on spatial neighborhood information and Fuzzy C-means algorithm is proposed. This algorithm will divide each image into 3x3 windows during the process of image segmentation, and incorporates the local spatial information and gray level information into FCM algorithm target function. In the iteration process, not calculate with single pixel , but in every neighborhood window data block, the purpose is to guarantee noise insensitiveness and image detail preservation. The new algorithm can overcome disadvantages of traditional and some improved FCM algorithm, such as noise sensitive, details information loss caused by filtering and ignoring of spatial neighborhood information .Experimental results indicate that the proposed method can get batter segmentation results and is more robust to noise.2. On the basis of point 1, using the outstanding global convergence of micro canonical annealing algorithm to optimize the improved FCM algorithm, and enable it to search for the optimal solution, reduce iteration and enhance the stability of the algorithm.3. A new method for handwritten digits recognition based on particle swarm optimization (PSO) and hidden markov model is proposed. This method defined 24 strokes with the sense of direction, so it can makes up the shortage which is sensitive for start-point in traditional methods, and reduces fuzzy brought by shakes. PSO optimize HMM, avoiding local infinitesimal obviously. Experimental results demonstrate that the proposed method can improve the recognition rate in most handwritten digits than traditional algorithms.