Research On Salient Region Detection Algorithms Based On Visual Attention

With the development of information technology, the size and complexity of image data continue to increase, which brings great difficulty to image processing and target analysis. In recent years, the salient region detection technology, represented by the visual attention, has become an important technical approach to improve the real-time and analysis accuracy performance of mass-data image processing.The traditional salient region detection model based on visual attention has some defects, such as poor contour detection and Mosaic results causing by neighboring interpolation. On the other hand, the spectral models that meet the real-time requirement also have some drawbacks such as lack of bionics meaning and unstable detection result. Responding to these problems, on the basis of visual attention, this paper analyzed the distribution characteristic of salient region in spatial domain and frequency domain and explored the saliency detection technology in several areas. The main work of this paper can be described into four aspects below:(1) Mechanisms of human visual attention were analyzed and we introduced the mechanisms in both areas of neurobiology and psychology. In addition, the classic visual attention model based on feature integration theory and the Koch neurobiological framework was discussed.(2) The structural similarity theory based on human high-level vision was studied, and we proposed two saliency detection methods of local and whole based on feature integration theory, these methods uses the SSIM and self-similarity operator to appraise the visual saliency. Then, the proposed methods were optimized. The experimental results show that our method overcomes the weakness of Mosaic causing by neighboring interpolation and it show better performance on contour detection and anti-noise.(3) The relationship between the spectral characteristics of the image and the saliency was discussed, and a saliency detection method based on phase spectrum and tuning amplitude spectrum was proposed. This method tries to change the spectrum distributions of the images to inhibit the redundant image features and enhance the salient image features. Then we introduced four spectrum tuning methods and discussed the relative strength of these methods. Meanwhile, we find the biological meaning of this spectrum method by studying the lateral inhibition mechanism of human vision. The experimental results indicated that our methods could not only detect the salient region of multiple images effectively, but also satisfy the requirements of real-time.(4) According to the clump target characteristics of the remote sensing image, two saliency detection methods that proposed above were improved and the experimental results indicated that our improved methods could detect the target region effectively. In summary, this paper studies the key technologies of human vision and proposed some salient region detection methods in spatial domain and frequency domain. The experiment results show the effectiveness of the proposed methods.