Texture Image Segmentation Method Based on Multi-layer CNN

抄録

This paper presents a new texture image segmentation method, which combines some texture feature images (the gray value of pixels in feature image represents the texture feature of the same pixels in texture image) into a binary value image that separates image into different texture regions. Based on the idea of separating images by edges between different texture fields, after obtaining texture feature images, we consider the texture image segmentation problem not as a pattern classification problem but several texture edges combination problems, which are simple binary value image processing problems like Edges extracting, Holes filling, Lines thinning and shorting. A new multi-layer cellular neural network (CNN) called MLCNN is proposed. Different with the standard CNN, in an MLCNN, Multiple templates can filter the input one by one, and each state value provides multiple outputs. Some discrete MLCNNs are designed for the binary value image processing problems mentioned.