Research On Visual Saliency Based On Deep Learning

In daily life,people are dealing with massive information.Faced with such amounts of information,the human eye has a unique attention mechanism to focus on the most important or interesting areas,which are called salient areas,and the related research is called visual saliency.Visual saliency model can quickly extract important regions during media processing tasks,and allocate limited computing resources to the locations we care about.It has important research value,especially in image video compression,target tracking,quality assessment,and pedestrian abnormal detection applications.In this paper,the existing image and video models and the saliency detection methods of video and high dynamic range images are analyzed and researched in detail.The research work mainly includes the following three points.1)In the study of high dynamic range images,this paper first studies the imaging technology of high dynamic range images,and combines the advantages of the existing excellent low dynamic range image saliency models to design significant images for high dynamic range images.Sexual testing framework.The framework extends the existing static image model into a high dynamic image model by means of multiple exposure decomposition.At the same time,the performance benchmark of the existing methods is established,and the performance of existing static models and high dynamic saliency is carefully compared.Experimental results on two open high dynamic range image eye movement datasets show that the proposed method has better performance than the existing high dynamic range image model.2)In order to compare the proposed video model with other existing methods,this paper builds an eye movement data collection platform,and organizes experiments to collect video eye movement data sets.Compared with the existing video dataset,the video source we selected is lossless HD video,which avoids the interference caused by the distortion caused by video compression.In addition,the selection of scenes is more abundant,including more and more challenging sports scenes.3)Based on the related research work of the existing dynamic video model,this paper proposes a video saliency detection model based on 3D convolutional network.The model accepts a separate video clip as input and improves processing efficiency compared to existing saliency models.In the existing public data set and the video eye movement data set proposed in this paper,a sufficient comparison experiment is carried out.The results show that the model has excellent performance in processing efficiency and prediction accuracy.