Research On Image Segmentation Techniques For Textile Images

Recently, the development of CAD and CAM provided a great progress of textile industry. Before dealing with the fabric by computer, we need to scan the fabric, and save it in computer as digital image, which were called textile image. Then the designers will edit the textile image according to the particular request.Before editing the textile image, we need to segment the image by the colors. It is an important task for it will affect the result of the textile industry. But the image segmentation algorithms used now aren't fit for the textile images which are disturbed with "texture noise". There are two characteristics in textile images: fewer dominant colors and disturbed with "texture noise". So in this thesis, we will propose two image segmentation algorithms for textile images according to these two characteristics.The first one is a technique based on region splitting. It can be described as three steps, First, the dominant colors in a textile image were extracted according to the observation of a user. Then the proposed method scanned through the image, which was segmented into a multi-scale non-uniform tree based on a context model and local color information, and the segmentation result was recursively refined at each scale. Finally a simple post-processing step was needed to remove all small connected components. These pixels were allocated to the majority color of their neighborhood.The second one is an approach joint context and multi-scale. At first, the dominant colors and a uniform texture region in the textile image were extracted based on a human visual system. Initially a crude segmentation was processed at the lowest scale. By gradually reducing the size of blocks, and the segmentation result was recursively refined at each scale based on a contextual model and Baysian theory. Experimental results showed that this algorithm can achieve better segmentation performance than the first one.