The Study Of Regional Saliency Evaluation Method Based On Eye Tracking Technology

In recent years, as the image retrieval technology continues to improve and processing capabilities of computer continue to increase, the fields of image processing and computer vision have been developed rapidly. As one of the most important procedures in computer vision, the extraction of local invariant features obtains more and more attention. A large number of experiments show that, as an effective tool of extracting image features, the local invariant feature can be applied to most applications of image processing. However, traditional methods of detecting local features are mainly based on some simple rules to select stable feature points. Because these rules are not universal, how to gain stable feature points is still a problem which urgently needs to be solved. This paper studied the algorithm of Scale Invariant Feature Transform (SIFT) and adopted unsupervised and supervised learning methods to improve traditional SIFT algorithm and to obtain more stable local feature points.In the framework of image retrieval based on the bag of words model, the quantity of SIFT feature points included by an image is often fairly large and features are not strong enough, so the efficiency and performance of image retrieval systems will be influenced. Based on properties of SIFT feature points and the principle of visual saliency, this paper proposed the local symmetry measure method for SIFT feature points. Also based on symmetry in the framework of image retrieval to increase the utilization efficiency of SIFT feature points, it embedded the filtering method and weighted strategy of SIFT feature points. In addition, this paper adopted a ranking framework based on learning and eye tracking experiments. Based on the mechanism of visual selective attention, this paper used a data-driven method of supervised learning to evaluate the stability of each candidate feature point and selected stable feature points. The results of image retrieval experiments on the Oxford Buildings Dataset show that selection strategies of SIFT feature points based on symmetry and learning to rank can effectively improve the performance of image retrieval. At last, through the eye tracking experiment, this thesis collected an eye movement dataset with good universality and strong usability for research.