Research And Implementation Of Saliency Detection Based On Visual Attention Mechanism

With the increasing number of images today, computerized image classification in high efficiency is urgently needed. In order to tackle the issue, saliency detection has been introduced as one of its solutions. Whereas, the currently available methods of saliency detection, which can merely detect the fixed saliency areas, fail to detect the objects with distinct salience degree by simply changing the parameter values.In this paper, we simulate the "super column" of human visual cortex area V1, and construct an "integrated column" model based on the feature integration theory. The Adjustable Saliency (AS) model being adopted to classify the images, we design the Image Classification system(IC system) by combining local features and visual salient coding on the basis of bag-of-words model. The main contributions can be summarized as follows:1. We stimulate an "integrated column" model. As a part of this model, the AS model can detect the salient object with distinct saliency value by changing the "PupilWeight" value, which similar with the way of human pupils adjust their focuses. Accordingly the AS algorithms are designed, and the computational complexity is analyzed respectively.2. According to the solution, we design and implement a prototype system named Image-Classification on the basis of the object-oriented software development approach.3. Based on the scenario of the project, we design the user cases to verify the detection ability of AS model and the classification ability of IC system. Experiments on the largest publicly available database (including 1000 images) show that our AS algorithm outperforms the existing 7 saliency detection methods. The IC system can significantly reduces the number of codesize on the MSRCv2 database.