Violence Detection And Face Recognition Based On Deep Learning Method

With the continuous improvement of the "safe city" development,public safety has gradually become a popular issue of concern.Video surveillance technology has also been used intensively.The traditional video surveillance system mainly provides recording and storing function,which is far from being able to meet the public expecta-tion of advanced technology.To establish the intelligent video surveillance system,the following key issues need to be resolved:(1)How to quickly detect abnormal behav-ior in the video,give alerts in a timely manner,and minimize missing and false report.(2)How to accurately recognize suspicious targets under non-optimal condition(Single Sample,Low Resolution).(3)How to ensure real-time and accuracy of the system in massive data conditions.In recent year,deep learning has achieved excellent results in many fields such as computer vision,speech recognition and natural language processing.This has brought new opportunities for the development of intelligent video analytics technology.This dissertation studies the issues mentioned above based on the deep learning method,the main contributions are as follows:].To solve the complexity of quick and accurate anomaly detection in video,es-pecially violence detection,we proposed a violence detection method relying on 3D convolutional neural networks.This approach utilizes a large number of tagged video data to carry out supervised learning by extending the traditional 2D convolution into 3D convolution to detect the motion in the video,and then uses the spatial and tempo-ral information to construct the deep neural network model to achieve the accuracy of violence detection.Due to the end-to-end learning characteristics of the CNN model,there is no need to design complex hand-crafted features to describe physical motion,therefore reducing the complexity of the task.The experimental result shows that the proposed method can effectively detect violent activities in a single scene and crowded situation.2.A two-stage voting approach based on KPCANet model is proposed to solve the single sample per person face recognition problem.This approach utilizes the un-supervised deep learning model KPCANet,which is trained on the face image patches without using additional training data,to extract facial features.The model ensures the robustness of the extracted features to illumination and occlusion,and eliminates the impact on the recognition when handling facial deformation.In addition,a voting scheme is used to combine the predicted value of each patch to get the final recogni-tion result.When simple voting is not matched,this dissertation suggests the two-stage voting method.By expanding the candidates set of each patch,different patch will be given different weights,which further enhances the accuracy of the final result.The experimental results show that the proposed approach have a superior result when ap?plying in the four widely-used face datasets.The algorithm is better than the generic learning methods,especially on the unrestricted face data LFW-a.the proposed method achieves about 15%higher accuracy than SVDL and LGR methods.3,A low-resolution face recognition solution based on the convolution neural net-work model is proposed to solve the problem that face images cannot be accurately identified due to low resolution in video surveillance systems.This dissertation pro-poses two models:the multi-resolution CNN model and the CNN model based on spa-tial pyramid pooling(SPP).(1)The multi-rcsolution CNN model is an improvement of the existing"two-step method".The low-resolution images are up-sampled by bicu-bic interpolation,and the images obtained by up-sampling will be combined with those high-resolution images as training samples.The CNN model can learn the consistent feature representations of both the high and low-resolution images,and then the cosine distance is used to measure the similarity of features,and finally produce the recogni-tion results.Experiments on the CMU PIE and Extended Yale B datasets show that the accuracy of the model is superior to other methods,and the accuracy is 2.5%～9.9%higher than the CMDA_BGF algorithm which has the highest recognition rate among the existing methods.(2)The SPP-based CNN model is an improvement of the "cross-space method".By adding the SPP layer to the CNN model,the model can output the constant dimension features for different sizes of input images.Finally,the recogni-tion result can be obtained by comparing the similarity between the gallery images and the probe one.The experiment shows that compared with the multi-resolution CNN model,this method can eliminate the need for up-sampling operation so as to simplify the process of image preprocessing,while maintaining high accuracy.In addition,it also reduces the number of mapping functions that need to be learned in the traditional"cross-space method".4.To solve the problem of bandwidth allocation in data transmission in surveil-lance systems and to meet the demand of rapid and accurate analysis of big data,a deep learning model based on "Sea-Cloud Synergy" architecture is proposed.The sea-sides train the local deep neural networks(DNN)using local data,which then can be used to analysis data quickly and give the real-time response.Those sea-sides assist cloud system training by uploading local models and small amounts of data.The cloud sys-tem uses these local models and data to build more complex deep learning model and to optimize the performance of the global model.Experiments on MNIST,Cifar-10 and LFW datasets show that the method of Sea-Cloud Synergy architecture effective]ly reduces the bandwidth consumption of data transmission,and also ensures the speed of the sea-sides and the accuracy of the cloud system.The above methods have been partially applied to the Prior Strategy project "Real-time Processing System of Massive Network Traffic Based on Sea-cloud Collaboration"(Grant No.XDA060112030)of the Chinese Academy of Science.