End-to-End Background Extraction Based On Deep Learning

With the development of Internet of Things technology,it is more and more easy for people to get the video and image resources from the network.Because it contains a lot of useless information and occupies a large amount of computer memory,in order to reduce the required memory space and improve people's work efficiency,it is particularly important to extract the effective features.Background extraction,as the basis of target recognition and detection,has attracted much attention.Based on this application environment,this paper studies the precise background extraction under complex background,the main contents are as follows:An improved SUBSENSE method structure is proposed to train data in order to extract background images from video sequences.As a pixel-level segmentation method,it relies on binary features and color information to detect pixel changes.This allows easier detection of foreground images with similar backgrounds,while ignoring most lighting changes.In addition,the internal parameters of the method can be dynamically adjusted by using the feedback loop at the pixel level without manual intervention.The dynamic saliency detection of video sequence is carried out by combining CNN network with LSTM network(long-term and short-term memory network).Firstly,the CNN network consists of two sub-networks,including the target network and the motion network,to complete the preliminary static saliency detection of images.Then DNN of LSTM network is added to detect the dynamic saliency of video.LSTM network has the advantages of fast training and detection,and can effectively avoid a limited number of training video over-fitting.Finally,input video frames,background images and dynamic saliency maps are fed into DNN network together.With the help of LSTM network model,dynamic information and background features of salient features can be easily learned,and finally accurate background extraction can be achieved.The network is trained with LEDOV video data set,and the performance of the network is verified and tested with DAVIS and FBMS video database.Through this method,the final saliency prediction and background extraction have achieved good results.