The Research Of Interested Objects Detection Based On Visual Attention

With the development of information technology, it’s a hot field to effectively locate the useful image data from large images and videos. It is an important methods to extract obtain useful information and neglect useless information based on visual attention model.Firstly, we introduce the related work about the all visual attention models, and analysis the advantages and disadvantages of other visual attention model. In this paper, we borrow the ideas from bottom-up visual attention, and add the contest-aware saliency based on top-down visual model. We also try to combine the mechanism of neural network feedback with our visual model.Secondly, we introduce the method of building the scene memorial visual model in detail. The visual attention model will extract the non-spatial features (color, intensity and orientation) and spatial features. With the interaction between dorsal and ventral pathway, the spatial features will modulate the non-spatial features. Using the central-surround algorithm, we can combine multi-features to generate the finally saliency map in the Gaussian pyramid.In the saliency map, the interested object will be selected and then transferred. In order to improve the reasonability of the saliency map, we add the detection of context-awareness in different scenes, then correct the final saliency map with the static and dynamic scene context features. With the method of combining spatial、non-spatial and scene context aware features, compared with other model, this attention model can solve the problem of irrational objects selecting with the signal feedback mechanism.Then, we choose four videos in different scenes and three static images to test our proposed model. The results show that the proposed model may be more biologically plausible and have better real-time and high the object extraction accuracy than other visual model, which can help the embedded intelligent vision system to find out the interested objects more accurately and reduce computational resources. In summary, the main contribution of the paper is as follow:we combine the different features between bottom-up and top-down visual model, and add the neural network signal feedback mechanism to modulate the non-spatial features. We generate the final saliency map with the spatial and non-spatial features in the Gaussian multi-scale space, and then using the context aware to correct the final saliency map. This improved methods can effectively keep the important information of the image. The result of object selection and shift in saliency map is also better than other traditional visual attention model. The scene memorial visual model is more reasonable than others. In the future, our related work can be used to intelligent tracking devices...