Texture Image Segmentation Based On Sparse Representation

Sparse representation and dictionary learning achieves good performance in image denoising, image reconstruction and face recognition. It uses sparse coefficients and reconstruction error as the criterion for the pattern classification. Also sparse representation provides a robust low-dimensional representation for high-dimensional features. SRC(Sparse Representation Classification) is an important classification algorithm introduced recently, which provides a new way for application based on sparse representation. The algorithm uses the training sample as the base elements of dictionary. Test samples can be represented through a linear combination of training samples, in other words, data samples from the same class can be linear represented by each other. Therefore, based on the theory of sparse representation above, we propose the sparse representation based texture segmentation. The method of sparse representation texture segmentation is to convert the image segmentation into the pixel classification. Generally, the method of sparse representation classification is based on block feature, which is difficult to accurately represent the texture information. In this paper we use Gabor filters for texture image and then extract the Gabor features instead of block feature. Gabor filters can achieve the texture information with different orientations and scales. The disadvantage of Gabor filters is that it makes the feature dimension too high. But sparse representation provides a low-dimensional representation for high-dimensional features.The main contributions can be summarized as follows:(1) We comprehensively introduce the theory about sparse representation and dictionary learning theory, and focus on the method of sparse classification and clustering, as well as the classic discriminative dictionary learning method which laid the theoretical foundation for later application.(2)A sparse classification based texture segmentation framework is proposed.The algorithm first uses Gabor filter for texture image, then chooses some pixels from each texture as the training sample and extracts their Gabor features to initialize the dictionary. Finally, we select each pixel from the test image as the test samples and calculate their Gabor features. Then the OMP algorithm is utilize to calculate the sparse coefficients to classify the pixels based on SRC criterion. Compared with the block feature, the experiment shows that through extracting effective texture features to the sparse classification based framework can greatly improve the accuracy of texture segmentation.(3) A dictionary learning is introduced to the texture segmentation. On the basis of sparse classification texture based segmentation framework, a discrimination learning dictionary is introduced. A discriminative dictionary learning is used after initializing the dictionary, which improves the representation and discrimination of dictionary. And then we calculate the sparse coefficients of test samples. Two criteria for the experiment by introducing a dictionary learning algorithms Experimental results show that the correct segmentation rate increased. Two discriminative dictionary learning algorithms are introduced to the experiment, which improve the accuracy of segmentation.(4) A sparse clustering based texture segmentation algorithm is proposed. The algorithm improves the sparse subspace clustering. Firstly, it selects some pixels from the image by using uniform random sampling technique. Then, it performs sparse subspace clustering to these pixels. Then the algorithm uses clustering result as the dictionary to classify the rest of the pixels. This algorithm effectively reduces computation time of sparse subspace clustering, which makes sparse subspace clustering in image segmentation possible.