Research On Human Silhouette Extraction Method In Static Images

The human silhouette extraction in static images refers to the segmentation of human silhouette from static images,it is widely used in computer vision for human behavior identification,background segmentation and replacement.The human silhouette extraction in static images faces great challenges,including the diversity of human body posture,the variety of clothing,the transformation of light and the complex background.In recent years,with the rapid development of deep learning,the method based on traditional feature extraction in image processing is gradually replaced by deep learning,and the convolutional neural network has a great advantage in image feature extraction.Therefore,it is of great significance to use convolutional neural network to extract human silhouette.The main research contents of this paper are as follows:1.Aiming at the problem that traditional feature extraction cannot precisely segment human silhouette,a method of human silhouette extraction based on deep learning is adopted.A specific convolutional neural network structure is designed with introducing the full convolution neural network,deconvolution and network in network in the model,the human silhouette extraction in static images at pixel level is realized.2.In order to improve the performance of the model,an improved method is proposed based on the constructed convolutional neural network.The original image is preprocessed by Gabor filter,before introduced into the convolutional neural network model for training.Using the combination of Gabor feature and the convolutional neural network,a more accurate human silhouette extraction has been achieved.3.The effectiveness of the methods proposed in this paper is verified by means of VOC2012 data set and Baidu human segmentation data set respectively.The improved model is applied to a video monitoring system with privacy protection function,a part of video in the CAVIAR monitoring video data set is selected for testing,and the results are analyzed.The experimental results show that:(1)the human silhouette extraction method based on deep learning realizes the fast and effective segmentation of human silhouette,which reflects the feasibility of using deep learning to carry out experiments;(2)the goodness of fit of the improved method test results is improved by 10.96%compared with the original model in the VOC2012 data set;(3)the test results on the Baidu data set show that compared with other existing methods,the improved method can reflect rationality and validity in terms of accuracy and processing speed;(4)the test results of the improved method on the CAVIAR data set provide the theoretical basis and improvement direction for the video monitoring application with high accuracy and synchronization requirements.