Feature Selection Algorithm Guided By Eye Tracking Data

Feature selection is an important part of the target recognition technology and plays important roles in image classification. Upon effective feature selection procedure which illuminates irrelevancy and redundancy of features, both the accuracy and the efficiency of image classification tasks can be significantly improved. Although most existing feature selection algorithms perform rather well, the cognition mechanisms of human vision is usually neglected in the feature selection procedure for image classification tasks. To address such an issue, the eye tracking technology is employed to explore the cognition process of human vision whose results are then used to guide the procedure of feature selection. The main work of this paper includes:Firstly, a coarse feature selection strategy is proposed based on a novel improved quantum genetic algorithm(QGA) and the eye tracking data. In this strategy, an eye tracker is employed to obtain eye tracking data. With the help of these data, regions of interest(ROI) from human perspective are identified to represent an image, from which seventy five low- level visual features are then extracted. For these features, an improved quantum genetic algorithm is proposed to perform coarse selection based on ROIs of an image, rather than the whole image itself, in terms of classification accuracy of a support vector machine. Experimental results show that the algorithm based on the improved QGA can explore better combination of features in accordance with human visual cognition mechanism than the traditional QGA does.Secondly, a hybrid feature selection algorithm combining the filter and wrapper type ones is proposed for fine selection of features upon coarse selection. This algorithm integrates the efficiency of the filter type relief F algorithm and the effectiveness of the wrapper type SVM-RFE algorithm. Meanwhile, the evaluation criterion for features is also improved in this algorithm. Different from the strategy in the coarse selection stage where the selection operations are performed in feature wise manner, the fine selection operations are performed in component wise manner. Experimental results demonstrate that,as compared with traditional selection algorithms, the proposed hybrid feature selection algorithm can be implemented more efficiently and features selected with this algorithm can achieve more accuracy in image classification. Experimental results also show that the eye tracking data are of great importance to improve the accuracy of image classification.