Research Of Salient Object Detection Based On Deep Learning For Video

Salient object detection is one of the tropical issues in computer vision that aims at acquiring salient objects or regions in images automatically by designing algorithms which imitates human visual system.Video salient object detection has a wide range of application value and guiding significance in video compression and coding,object tracking and edge detection,which spark the enthusiasm of researchers in this field.In resents years,the saliency detection has made great progress with the advances in deep learning.In this study,we compared the methods for saliency detection and present a new strategy based on deep learning methods,which has a competitive performance in efficiency and accuracy.The main research contents of this thesis are as follows:Firstly,we summarized the research status of salient object detection in images and videos and reviewed the theoretical basis of saliency detection.We highlighted and analyzed four mainstream saliency detection models which provides the methodological basis and theoretical basis for further research on salient object detection.Secondly,inspired by the Generative Adversarial Networks and Elastic Net,we proposed a method for video saliency detection based on the conditional Generative Adversarial Networks and multi-paradigm weighted regularization.This method consists of two confrontational parts which are generating model and discriminant model.Through the adversarial learning and multi-paradigm constraint on the data set,the final model can finally well grasp the spatial-temporal features of video to be detected and complete the task of video salient object detection.Experiments on the public data sets showed that this method has a better performance and can meet the real-time requirements.Finally,in order to solve the problems of large inter-frame loss and insufficient learning of temporal features in video saliency detection,we combined the cyclic learning rate theory and the residual network and U-net network and proposed a salient object detection method for video that enhances temporal features.The method includes two consecutive stages.In the first stage,the model is trained through static images and video data sets.In order to enhance the temporal features of the video sequence learned by the overall model,in the second stage,the model is further trained by combining the pre-and post-frame information of the video frame with the saliency map of the model output in the first stage and then the final video saliency detection model is obtained.Experiments on the public data sets showed that the detection accuracy of this method has been further improved.