Research On Extraction Of Linear Structures And Segmentation Of Regions In Images

In images, linear structures usually refers to the contours used to define the target shape or the boundaries used to demarcate the regions, they provide image analysis and understanding a simple and reliable expression of the shape characteristics. Another technology which is closely related with extraction of linear structures is segmentation of regions, which refers to the seperation of those regions have different statistical properties in image. On the one hand, it can provide not only the structural information by the contours or boundaries constitute the regions, on the other hand, as in each segmented region, pixels hold the similar statistical characteristics, thus each region can be approximately viewed as a whole, through the feature extraction for each region, it also provides image analysis and understanding a simple and reliable expression of the regional characteristics.At present, in the fields of image processing and computational vision, researches on extraction of linear structures and segmentation of regions though have made significant progress, yet both still have no thorough solutions, both are still open topics. As both can provide the high-level visual processing, such as object recognition and scene classification, with important features of expression, they are always two important research topics.For extraction of linear structures, this thesis uses small straight-line segment as the compositional token, and models the linear structures under the framework of Marked Point Process. The advantage of this model is that we can simultaneously complete the detection and spatial grouping of tokens during the optimization. Experiments on both synthetic and real images showed that our method can capture most of the salient structures, while reduce largely the distracting edge elements due to texture or cluttered background.For segmentation of regions, this thesis casts it as the problem of clustering feature vectors of pixels. Specially, a robust clustering algorithm for feature space is proposed. Experiments show that our clustering algorithm could obtain more coherent results at the situation when the distribution of data is complicated, than some classical clustering algorithms. Due to the robustness of our clustering algorithm, only cluster the pixels in the L*a*b* color space, we can obtain the satisfactory segmentation result. In addition, for each image, in order to obtain a satisfactory segmentation results, how to automatically set the value of tuning parameters, is always a problem for many image segmentation methods in practical applications. Based on the theory of Minimum Description Length, this problem is put forward a solution. To verify the validity of our region segmentation method, experiments are performed on the Berkeley segmentation database, and the quantitative comparison with state-of-the-art segmentation methods is reported.Quantitatively evaluating the performance of segmentation is also important in the field of computational vision, however, there is still no standard performance measure. This thesis presents a new evaluation method for image segmentation results. For each pixel, we give it a weight by defining its degree of perceptual consistency among the hand-labeled segmentation ground-truths. The higher the degree of perceptual consistency, the higher the weight of corresponding pixel, our final evaluation index is calculated based on Jaccard Index with the weight map. Experiments show that, our evaluation index could better reflect the understanding of human visual perception on image segmentation from the ground-truth.Extraction of salient regions from image is important for many high-level vision applications, such as object recognition, content-based image retrieval. We propose a simple and effective model to compute the saliency for pixels in image. With the resulted saliency map for image, salient regions are extracted by using hysteresis thresholding. At last, we evaluate our method on a database which contains 1000 pairs of image and groundtruth.At last, we summarize the presented work. According to the imperfect aspects, we analyze and discuss the future work.