Research On Visual Attention Mechanism And Its Application In Human Detection

The mechanism of attention selection is one of the important characteristics ofhuman visual perception. Visual attention has been studied by researchers inphysiology, psychology, neural systems, and computer vision for a long time. It isuseful for many computer vision tasks such as content-based image retrieval,segmentation, and object detection. To probe the method of computationally modelingthe mechanism of visual attention and its application, the thesis focuses on thefollowing aspects:First, the status of research on bottom-up visual attention model and top-downvisual attention model is reviewed. The attention model based application in imagesegmentation, object detection are also introduced. Visual saliency map is better thanoriginal image for image segmentation and object detection from further analysis onthe mechanism of visual attention. Image segmentation based visual saliency map andobject detection based on saliency map and predicting fixation are useful for imageunderstanding.Then, a visual attention model with joint spatial and domain information isapplied on content-based object detection. Our method combined bottom-up saliencydetection method with traditional image segment method (mean-shift) and comparedthe regional saliency with global image saliency to detect salient object. We haveevaluated the results of our approach on the publicly available database. To the best ofour knowledge, the database is the largest of its kind, and has ground truth in the formof accurate human-marked labels for salient regions. The resulting saliency maps arebetter suited to salient object segmentation, demonstrating either higher precision orbetter recall than the six state-of-the-art techniques. We also use our method to detectpeople in a well-known database, result shown that it can get a better subjectiveeffect.Finally, the model with joint spatial and domain information is applied to peopledetection combined with top-down object information and context information. Thepaper uses the model with all the three resources to generate artifact fixationsequences. The result demonstrates that top-down information is the most importfactors. A comparison between artifact fixation sequences and ground truth indicatesthe efficiency of the model consisting of three factors with more similarity and higher accuracy.