Research On Superpixel Based Boundary Detection And Segmentation Algorithm Using Pointwise Mutual Information

Boundary detection and image segmentation algorithms are the foundations of image analysis and image recognition which become more and more important and highly concerned in image processing. These skills have been used widely in military, medical, agriculture, industry and the Internet fields. Point-wise Mutual Information(PMI) measures the similarity between two different pixels which providing a new feature representation method for boundary detection and segmentation algorithm by using the conception of probability theory and mathematical statics and it has been received widely attention. The prime target of this paper is to enhance the PMI feature extraction by using the middle structure information provided by super-pixel segmentation and to improve the unsupervised edge detection and segmentation algorithm. The main contributions of this paper are summarized in the following:(1) Super-pixel Point-wise Mutual Information Feature. This paper extracted the SPMI into global features and combined the super-pixel proposed a new feature representation which obtained the connection between basic features of each pixel and the middle structure information of images which called Super-pixel Point-wise Mutual Information(SPMI). The structure information provided by super-pixel segmentation would be used to limit the position of the samples as the conditions of sampling process which could make the sampling procedure more purposeful and targeted, and would extracted the features of images as many as possible even thought the number of sampling points had been reduced.(2) PMI boundary detection algorithm based on super-pixel. This paper overcame the disadvantage about randomness and insignificance of partial sampling points by using the Super-pixel Point-wise Mutual Information feature to guide the sampling process. This algorithm could reduce the redundant points and increase the proportion of effective sampling points which could improve the accuracy of the algorithm. Through the comparison between the original boundary detection algorithm and developed one, the experimental results demonstrate that the super-pixels based boundary detection algorithm enhance the accuracy in unit time. The average precision of SPMI algorithm based on super-pixel method is 0.7917 when the number of sampling points is 3500, while the accuracy rate is 0.7915 of the original detection by calculating 6000 sampling points.(3) Spectral clustering segmentation algorithm using Super-pixel Pointwise Mutual Information. This algorithm used the SPMI feature and the Ncut spectral clustering method to complete the whole image segmentation process. In this method, the characteristic values of the pixels in SPMI feature domain and the geometric feature domain would be calculated and the different features of the current two scales would be fused to construct the affinity matrix. Secondly, the characteristic system of the affinity matrix which obtaining the direction and the SPMI was calculated. Finally, by using the method of 2-way decomposition(2-way Ncut) which was performed on the affinity matrix the algorithm would obtain the segmentation results. Experimental results show that the accuracy of the proposed algorithm is improved by 23% comparing with the original Ncut algorithm and also improved 15% comparing the supervised LFPA algorithm based on the Ncut criterion.