Research Of Object Detection Algorithm Based On Visual Attention Mechanism

Visual attention mechanism is the selection mechanism of huma n visual, allowing humans to handle complex scenarios quickly to get the wanted content. Object detection is fundamental and hot topic of computer vision, the research of object detection based on visual attention mechanism has great significance. Saliency detection algorithm is a kind of object detection algorithm that based on visual attention mechanism by detecting the most significant area in the scene that can attract the attention of human visual to complete salient object detection. Many excellent algorithms have been proposed since L. Itti put forward saliency detection in 1998. However, many saliency detection algorithms only extract single feature or just fuse different features with average weights. To solve above problems, this paper proposes a multiple features multi- nuclear fusion saliency detection algorithm, which extract include color, orientation and compactness three kinds of features to compute saliency maps and fuse saliency maps with adaptive weights. The main research content of this thesis is as follows:1) The SLIC(simple linear iterative clustering) superpixels segmentation algorithm is been researched. It has been shown by a lot of research results that region-based saliency detection method is better than the pixel-based method. The SLIC superpixels segmentation algorithm can over-segment images into superpixel regions that boundaries adherence as well as compactness are well, which is used by many excellent algorithms for image pre-processing. Based on the segmentation results of SLIC algorithm, a graph model is constructed to propagate the similarity between regions to compute compactness saliency map.2) The computation of multiple features include color, orientation, and compactness saliency maps has been researched, and the results of three kinds of saliency maps are analyzed.3) An adaptive weights distribution method based on the spatial variance of different feature saliency maps is proposed. Compared with a lot of algorithm that fuse different saliency maps with average weights, our fusing method can assign adaptive weights to different saliency maps, which is more robust and the performance is better. Up-sampling algorithm is researched to assign the saliency value to each pixel for generating full resolution saliency map.Experiment on MSRA-1000 data set is to verify the effectiveness of the improvement of the proposed method. The results of performance comparison experiment on MSRA-1000, THUS10000 and ECSSD data sets show that the proposed algorithm outperforms the other five state-of-art algorithms include FT, SF, LR, HS, and GBMR.