Image Feature Selection Research Combined Eye Tracking Data

With the rapid development of the Intern technology, more and more high-dimensional data, especially the image data, need to be addressed in many fields. In this process, feature selection, which selects an optimal subset from the original feature set, is an important step and has become a hot topic in the field of pattern recognition. By removing redundant and irrelevant features, effective feature selection can reduce the feature dimension and improve the classification efficiency. While most existed methods have achieved promising better performance with mathematically sound algorithms, the functions of humans themselves have seldom been taken into account in the process of feature selection. Feature selection process with human beings involved is of particular importance in image recognition. In this thesis, the effects of human beings in feature selection are considered in terms of the eye tracking data. Eye tracking data here are introduced to assist selecting regions of interest from which features are extracted. A new idea that combines Filter and Wrapper models is proposed for selecting image features and to achieve a balance between the efficiency and classification performance. The main contributions of the dissertation are as follows:Firstly, a rough selection algorithm called RS-E is proposed, in which eye tracking data are employed for image feature selection. Problems of "semantic gap" exist for classification algorithms based on low-level features such as color, texture, shape and so on. To tackle such a problem, eye tracking data are utilized to select human"s regions of interest in images, from the most distinguishing areas from which some effective features are selected. The employment of eye tracking data is helpful to simulate the process of human when recognizing objects. Experimental results show that both the performance of feature selection and the efficiency of classifier are greatly improved after introducing the eye tracking data.Secondly, based on the complementarity of the Filter model and the Wrapper model, an algorithm named mRMR-SVM-RFE is proposed to balance the efficiency of feature selection and the accuracy of classification results. As a composite algorithm of the Filter and Wrapper method, this algorithm is used for fine selection of image feature in combination with the RS-E algorithm. Experimental results show that, as compared with several traditional algorithms, higher image classification accuracies are achieved with the selected subset of features from the proposed algorithm. Additionally, an improved mRMR-SVM-RFE algorithm is also proposed, in which both the mRMR criterion and the SVM-RFE ranking criterion are taken into account in the improved evaluation criteria. Experimental results show further improvement of the performance in image classification.