Research And Implementation Of Monitoring Video Quality Diagnosis Method Based On Transfer Learning

Nowadays,security has become increasingly important.In classrooms,campuses,shopping malls,banks,roads,etc.,surveillance devices can be seen everywhere,and every frame of surveillance images contains important information.When transmitting a large amount of video data,various image distortions may occur,ultimately leading to the loss of important monitoring information.Therefore,feedback on monitoring quality is particularly important.The deep neural network model requires large-scale video data,and the server hardware required for network training and learning is expensive and time-consuming.In addition,there is a lack of real-time application inference support for neural networks in video capture on-site quality diagnosis.In view of the above problems that need to be solved urgently,this paper uses the no reference quality evaluation method to establish a small monitoring video fault dataset,and uses transfer learning to identify monitoring faults.The main research achievements are as follows:(1)Aiming at six typical fault types,such as black screen,stripe,snowflake,occlusion and blur,which are common in monitoring at present,a method of monitoring video quality detection based on transfer learning is proposed.Establish a monitoring fault data set,enrich the data set by cutting and rotating,and then calibrate the 4000 chapters of pictures one by one,and convert them into the format needed for network detection,and judge whether there is any quality problem through image detection.(2)By comparing the current classic object detection networks,the YOLOv5 object detection network was ultimately selected.The transfer learning method is used to reduce the requirements of the model on a large amount of data,and optimize the network structure for the occlusion,image blur and other fault types.By adding the attention mechanism CBAM to the C3 module,the model strengthens the attention to the target information,and optimizes the network parameters,which significantly improves the detection accuracy of the network model.(3)YOLOv5 network model based on transfer learning,in order to improve the robustness of the model,the original spatial pyramid pooling is replaced by a more complex SPPFCSFC,which enhances the fusion of image local features and global features,and improves the expression ability of feature maps.And replace all convolution operations in the Neck module with GSconv to reduce the complexity of the model.(4)Finally,in order to make the model have good ability in practical application,the trained model is deployed on NVIDIA Jetson Nano embedded platform,and the information is read by external cameras,and then the inference ability of network model is accelerated by Tensor RT and tested by FP16.The final test frame rate is 13 FPS,and the precision loss is less.In this paper,based on the research of fault detection of surveillance video,aiming at the problem of limited data set,a fault identification method of surveillance video is proposed,and it is improved based on YOLOv5 network model.Compared with the basic network,the average detection accuracy is improved from 91.4% to 95.3%.The experimental results show that the new model studied in this paper has better detection ability and provides new technical support for video fault diagnosis.