Research And Application Of Visual Attention Model With Structural Complexity Based Feature Fusion

Visual attention is the ability that human focuses attention on a few regions of interest rapidly to process them at high priority with limited resources, when facing a complex scene. Researching on computational modeling of visual attention is not only helpful to understand the underlying mechanism of human visual system, but also have important application in image analysis and understanding, object detection, information retrieval, robot vision, etc. Existing bottom-up attention models include spatial-domain models and frequency-domain models. Spatial-domain models are biologically plausible, but they suffer from high computational complexity. Frequency-domain models often run faster, but they have no biological evidence. Following these models, we propose a visual attention model with structural complexity based feature fusion, and then optimize it on processing speed using Haar-like features instead of Difference of Gaussians features. Moreover, we apply the proposed model to image object segmentation. The main contributions of this paper can be described as follows:1. A visual attention model with structural complexity based feature fusion is proposed. According to the knowledge on human vision system, the input image is first converted into perceptual color space LAB from physical color space RGB. Then multi-scale features for each color component are extracted using Difference of Gaussians. At last, all feature maps on different scales are fused into the master saliency map according to their structural complexities. Experimental results show that overall the proposed model outperforms the six state-of-the-art models, and has better consistency with psychological experiment results.2. Considering the high computational complexity of our proposed model, we further propose a fast visual attention model based on Haar-like features with the same structural complexity based feature fusion method. At first, we use integral images to fast compute multi-scale Haar-like features for each color component. Then all feature maps on different scales are fused into the master saliency map according to their structural complexities. Experiment results show that under the premise of little impact on performance, Haar-like features can greatly improve the running speed of our model.3. In order to segment out the salient object automatically, we combine the proposed visual attention model with the image segmentation algorithm Graph Cut. For a given image, we first compute the saliency map using visual attention model and get seeds for salient object through thresholding. Then we perform iterative segmentation with Graph Cut to grow the seeds, and thus segment out the salient object. Experiment results show that compared with the other three typical visual attention models, our model can provide the seeds for object segmentation more effectively to obtain better results.