Research On Natural Scene Image Salient Regions Detecting Algorithms

When facing a complex scene, the human visual system can easily and efficiently detect and extract information interested. Those interested information have the priority to be delivered to human sense. It is necessary to research on saliency detection algorithms to simulate human selecting visual system in study of computer vision. Salient regions in images refer to the parts best represent the present image contents and those that human attention most attracted in. There are some common problems existing in current algorithms. First, most models are not able to highlight the whole salient regions uniformly. Second, most models don’t have good robustness in complex background. Third, most models have computational redundancy because of the unknown size of salient objects. In order to solve these issues, this paper presents two models of detecting salient regions in natural scene images.The first model proposed is simultaneous sparse coding and multi-scale based salient regions detecting algorithm. First, we construct multi-scale Gaussian pyramid for input images to obtain multi-scale features. Then, we apply the simultaneous sparse coding framework to compute the sparse coding coefficients of every image patch and compute the initial saliency map. Last, we fuse the multi-scale saliency maps to achieve the final result. Simulation experiments results prove that our algorithm is effective and reliable.The second model proposed is super-pixel and global contrast based salient regions detecting algorithm. First, we apply the SLIC super-pixel segmentation method for input images. Then, we calculate the global color contrast of every super-pixel and give each contrast a certain similarity weight related to the space distance. Last, we utilize the local neighborhood similarity constraints to refine the preferred contrasts and achieve the final result. Simulation experiments results prove that our algorithm is effective and reliable.The paper also introduces two applications. The first is texture information based classification of high-resolution SAR image. The main idea is to utilize GLCM to analyze image texture, select appropriate features by Bhattacharyya distance and last perform the final classification. The second is non-local neighborhood based multi-temporal SAR image change detection. The main idea is to use non-local meansto operate every image patch and then compute the difference value of each pixel.