Saliency Detection Research Based On Learning

Saliency detection is one of the most important research fields in computer vision.It is widely used in image segmentation,image compression,target tracking and image recognition.Traditional saliency detection algorithms rely on the underlying features of image data,or use a large number of manually labeled truth map training machine learning model,with large calculation and slow calculation speed,it is difficult to deal with real-time changes in the actual environment.The human visual system has the visual attention mechanism to locate the target area quickly and efficiently.Therefore,this paper simulates the human visual system,and proposes an active vision framework based on gaze.The main work is as follows:(1)This paper proposes a saliency detection algorithm framework based on learning,which is divided into four parts: gaze point prediction,attention area division,classification model training and pixel classification.gaze point is predicted by the improved spectral residual algorithm.using large value points to divide the region of interest and sampling;training the classification model with the extreme learning machine;in pixel classification stage,using the classification model to scan the scene to realize the saliency target detection.(2)Two strategies for attention area division are proposed.1.Draw a local small box with the maximum information entropy around the center of a single large value point;2.Draw a large global box with multiple significant points as boundaries.The result of the division of the gaze region directly determines the quality of the sampled samples and indirectly influences the performance of the model,it is an important step of the algorithm.(3)An adaptive image scacle selection algorithm is proposed.In order to solve the problem that the frequency domain analysis algorithm is sensitive to the change of image scale,this paper chooses several large points in the prediction map of the gaze point and computes their dispersion in the airspace,and determines the optimal scale according to the size of the dispersion.The proposed algorithm is tested on six published standard datasets and compared with other classic saliency detection algorithms.Experimental results show that the proposed algorithm can quickly and efficiently highlight the target area while suppressing the perimeter noise effectively.Compared with the classical algorithm,the proposed algorithm has higher accuracy rate,recall rate and F-score compared with the classical algorithm,and has less running time,It is possible to extend the significance detection model to practical application.