Efficient Scene Classification Based on Maximum Entropy Policy and Visual Attention

Through the study of attention selection mechanism based on Amplitude Modulation Fourier Transform, a novel scene classification method based on maximum entropy policy and visual attention is proposed in this paper. This method adopts Amplitude Modulation Fourier Transform to construct saliency map, and adaptive Gaussian filter is used on the old saliency map to get the new information-rich saliency map. Then reordering the maximal saliency points according to entropy in the neighborhood of maximum points, top-ranking maximum points are considered as the center of region-of-interest. Eigenvectors representation algorithm based on histograms of oriented gradients is designed to improve the separability of region-of-interest, finally scenes matching are implemented through calculating Euclidean distance between eigenvectors. Compared with traditional methods, it has a good invariance in image scaling, rotation, translation and robust across a substantial range of affine distortion, meanwhile having better real-time. The experimental results demonstrate that the method is well applied to scene classification and retrieval with better time-consuming.