Visual Saliency Detection Based On Objectness Proposals

With the rapid development of Internet technology and the popularity of electronic products, images have become an important information carrier. How to extract the useful information from images becomes the key factor restricting the development of computer vision. Saliency detection of images can effectively distinguish the attention and non-attention area. By obtaining the attention area, feature extraction,image enhancement, image edge detection and image recognition and understanding can also be achieved. Through the saliency detection model, the region with high degree of attention in the image can be obtained, which can be used to separate the significant area from background, greatly improving the image processing speed.In this paper, four representative models of saliency detection are introduced, and their effects are compared. Through the study of Ground-truth map, it is found that the region where objects exist has higher significance. Therefore, this paper presents a new multi-level significance detection model based on the objectness proposals theory. In this model, we use the multi-scale super-pixel segmentation, feature extraction based on region contrast and encouraging diversity ranking mechanism to get the significance region with clear edge and complete content. Firstly, the SLIC method is used to segment the image, and calculate the boundary probability of each cluster center. Then, the similarity between different regions of the image is obtained by the color contrast, texture contrast and spatial distance. Next, by merging the similarity, we get the proposal regions. Finally, encouraging diversity ranking mechanism is used to obtain the region of high target probability,and the final saliency map is got by normalization.MSRA, MIT300 and PASCAL-1500 image datasets were selected to verify this model’s detection effect quantitatively and qualitatively.And we also compare this model with other representative models. The experimental results show that this model is better than the existing models in terms of precision rate, recall rate, and PR curve, and prove the effectiveness of this saliency detection model based on object proposals.