Research On Visual Attention Mechanism With Auxiliary Information

Visual attention mechanism is from the biological mechanism of retina.When we observe complex scenarios with multiple targets,the retina will pay attention to a small number of targets and allocate limited processing resources to them in every moment.Because of the selectivity and focusing capability,the visual attention mechanism is widely used in the field of computer visual.Automatic image annotation is a typical application in computer vision field,which annotates specified image automatically and yield to the tags of the image.As the images in natural scene are almost with multiple labels,the automatic annotation for multi-label images is more useful.The multi-label image annotation system yields to a tag in every iteration and the selection of features and the distribution of processing resources are important to the final annotation results.In this paper,the results of multi-label image annotation is the regarded as the quantitative index.The visual attention model is built with the deep neural network and we improve it with some methods.There are two problems of visual attention mechanism,which is the insufficient differentiation capacity and focusing capability.Differentiation ability is the ability that the attention model pay attention to the different objects in different time.The attention model focus on one object in the previous time and we want to pay attention to another object in the next moment and minimize the interference of the object in the previous time.In this paper,we add the history information to the attention model in order to improve the differentiation ability,which inputs the results of the previous iterations to the attention model in the current iteration and reduce the history information in some rules to get more accurate results.Focusing capability is the ability the attention model pay attention to one object,which is we want to distribute the processing resources to one object and ignore other objects and clutter background.In this paper,we introducing a significant information to increase the focusing ability of the attention model.We extract the underlying features of images and contrast the features of different pixels to get the significant information and add it to the attention model.When the attention model select the attention region referring to the significant information and get more accurate results.In the paper,we mainly study on the visual attention model based on the deep neural network.We improve the differentiation capacity and focusing capability of attention model by adding the history information and significant information.There are a lot of experiments to prove the effectiveness of the proposed approach.