Visual Attention Model And Its Application On Object Detection

In the information society of data soaring, image data processing has become an increasingly significant problem. While the image data needed to be processed is only a tiny part of the whole image data set in actual problems. Therefore, it is of great significance for introducing visual attention model based on human visual attention mechanism into image processing such as object detection.In this paper, visual attention model and its application on object detection were studied. The principal tasks as follows:(1) We proposed a spectral residual visual attention model based on discrete cosine transform. It is an improved visual attention model which can overcome the defect of Hou’s method. In the model, different features such as color, intensity and direction are extracted according to different images. After this, conspicuity maps are obtained by using the spectral residual based on discrete cosine transform. All conspicuity maps are combined nonlinearly to obtain saliency map for guiding attention and detecting regions of interest. The focuses of attention are shifted by mask. The results show that the proposed model can obtain saliency map and detect regions of interest better than Hou’s method.(2) We proposed methods based on visual attention model for object detection of static images and image sequences. These methods use the significance of saliency map and shift rule of focus obtained from visual attention model in the process of detecting. Experiment results demonstrate that the methods can detect object effectively.(3) We proposed a visual attention model based on dyadic wavelet transform. In this model, the wavelet decomposition of each early visual feature is performed and the modulus of dyadic wavelet transform corresponding to each feature is obtained. Wavelet decomposition level of feature is one here. Then redundant information included in the modulus of wavelet transform is removed to obtain conspicuity map. The saliency map is obtained by combining all conspicuity maps nonlinearly. This model solves the problem existed in spectral residual visual attention model based on discrete cosine transform.