Studys On Structure Extraction Methods For Texture Images

Structure extraction efficiently from texture images, such as natural images with multiple level textures, or ―main structure & texture‖ mosaic or graffiti, is the basic research subject of computational photography and image analysis. It can not only greatly improve the quality of image understanding, but also be applied to object detection and saliency analysis as well as other computer vision tasks, having great research and application value. The diversity and complexity of texture brings some difficulties, making the structure extraction work be a challenge. The current study focusing on the prior measure of distinguishing structure from texture, which often required regular texture, cannot get satisfactory results of images with universal or random texture. Meanwhile, the efficiency of the current algorithm has yet to be improved.In this thesis, firstly we apply the new prior, named Hyper-Laplacian gradient prior which could describe structure feature, to the basic model, proposing Hyper-Laplacian gradient prior-based image smoothing model(Lp G) from the perspective of the optimization model. The model can effectively and efficiently maintain primary significant structure and remove small extraneous details, getting a good application on the cartoon artifacts removal and tongue image segmentation. However, the results of salient texture images are not so good. Then, we propose a blur guidance relative total variation method(blur_RTV), applying the more correct guidance image with Gaussian blur to the original relative total variation model. The new RTV method can obviously enhance the degree of removing texture at structural edge and ensure the smoothness of structure, which fit for such images with single-direction prominent texture.The above two methods significantly improve effects at some extent, but they all require priors or texture pattern. Therefore, in order to improve the universality, this paper innovatively studies structure extraction method from the viewpoint of pattern classification. We estimate the structural and textural attributes of each pixel to obtain a structural contour feature map, and then combine it with the weighted total variation, forming the structural contour feature learning based weighted TV method(LTV). Training on large dataset ensures the robustness of our method, so LTV can gain satisfactory results on various textures images, extracting complete structure information and removing texture at maximum extent. We compare our methods proposed in this paper with two state-of-the-art algorithms in detail, providing comprehensive summarization on the computational efficiency, the applicable situations, main characteristics and limitations of each algorithm. In the end, we will put more effort into feature learning of LTV method because of the discontinuity at structural edge in future work.