Research On Pooling Strategies For Image Segmentation Quality Evaluation Technology

Image segmentation is a basic and key technology in the field of computer vision and image processing. Although many segmentation algorithms have been proposed for a variety of specific applications, but, so far, there is still no unified segmentation system. The existing theories and methods of image segmentation are various, and many new methods are constantly being proposed. So, how to choose the appropriate segmentation algorithm is the core issue that segmentation algorithm designer would meet in a long term. Therefore, it is very necessary and significant to evaluate the different segmentation algorithms. In order to simulate the segmentation evaluation ability of human visual system, and to take advantage of the influence of prior knowledge internal structure relation on the segmentation quality structural characteristics, the main content of this thesis is about the image segmentation evaluation technology based on pooling strategy.Firstly, the background and research status of image segmentation and image segmentation evaluation are introduced. Followed by a review of the classification, research contents and evaluation criteria of image segmentation evaluation methods. Then, the database used in this thesis, including the Berkeley natural image database and the Image segmentation quality evaluation database set up by our own laboratory is introduced. Moreover, three kinds of supervised image segmentation evaluation algorithms, i.e. Probabilistic Rand Index(PRI), Variation of Information(VOI), Global Consistency Error(GCE) are introduced, and the performances of each algorithm are also given. Furthermore, two improved evaluation algorithms are proposed by combining visual saliency and segmentation quality into original evaluation algorithm. After that, two kinds of experimental methods including the subjective evaluation and the meta-measures are applied to evaluate the improved evaluation algorithms. These two experimental results all show that the new image segmentation evaluation technology proposed improves the accuracy of segmentation evaluation in comparison to the original methods. Finally, a software system for image segmentation evaluation is designed and implemented. The basic function of the system is complete, the interface is friendly and simple, and the operation is convenient.