Classification Model Of Visual Attention Based On Eye Movement Data

Visual attention is a very important part of the human visual system. Most of the existing visual attention models emphasize bottom-up attention, considering less top-down semantic. There is few specific attention models for different categories of images. Eye tracking technology can capture the focus of attention objectively and accurately, but the application of visual attention model is still relatively rare. Therefore, we propose a classification model of visual attention (CMVA) combined bottom-up with top-down factors, which train classification models for different categories of images on the basis of eye movement data so as to predict visual saliency.The main contributions of this thesis are as follows:1. In order to train visual attention models for different categories of images, we selected four typical types of images that have a higher frequency from several image libraries and network images. Eye movement experiments were performed on each categories of image, the focus map of the experimental results was put Gauss convolution to get the "ground truth" saliency map to show the region of the human eyes, and construct the eye movement data set of the four kinds of images.2. We selected positive and negative samples from the "ground truth" saliency map of each categories of image and extracted different high-level features from different categories of images. The main difference is target detection feature. At the same time, a series of low-level features were extracted to define the significant position, and a linear support vector machine was used to train the classification model of visual attention CMVA. Finally, three evaluation indexes were compared with the other eight models. To make unknown class image automatically select the corresponding category of visual attention model, we set up four types of image as the training set and marked the four categories, namely,1,2,3,4. Then part of the low-level features and the high-level features were extracted, and the SVM was used to train an image classifier, so that the unknown image could be obtained the category to which it belongs, and then used this kind of visual attention model.3. There were some errors in eye movement experiments conducted by the eye tracker. In order to compare eye movement training model and real saliency map training model, subjects manually labeled the region of interest four kinds of images, to get the manual "ground truth" saliency map. Finally, the eye movement model and the manual model were compared by the evaluation index.Our model was compared with other existing eight models, proving its superior performance than other models. Compared with the manual model, the performance difference is little, but it can replace the heavy manual marking, which greatly improves the efficiency, and has a great value of research.