| In order to measure the temporal quality,spatial quality and comprehensive quality of videos transmitted by Visible Light Communication(VLC)accurately,the methods of quality assessments were studied and this paper was entitled "Technology of Quality Assessment for Videos Transmitted by VLC Based on Multi-Task Deep Learning ".In this paper,framework of video quality assessment based on multi-task deep learning is designed.Then the technologies of video temporal detection based on object detection with multi-level features,video spatial detection based on Siamese Network,video quality assessment based on multitask deep learning are studied.At last,experiments of temporal detection,spatial detection and video quality assessment are executed.The research was supported by a major special project for collaborative innovation of industry,university and research in Guangzhou(No.201604046005).The research of temporal detection,spatial detection,video quality assessment and multitask learning at home and abroad are analyzed in this paper.The main points of this paper are as follow.(1)Design of framework about quality assessment for videos transmitted by VLC based on multi-task deep learning.In this part,the requirements of quality assessment for videos transmitted by VLC are analyzed,the scheme and process of quality assessment are also designed.The implementation of key technologies is focused on,which contain the technologies of video temporal detection based on object detection with multi-level features,video spatial detection based on Siamese Network,video quality assessment based on multitask deep learning.(2)Research on the technology of video temporal detection based on object detection with multi-level features.In this part the object detection network with multi-level features based on Single Shot Multi Box Detector(SSD)is built.The structure of SSD object detection network is optimized by using Feature Pyramid Networks(FPN)to improve the detection rate and accuracy of SSD object detection network for small object.The video label information is extracted and verified according to the object detection results to achieve the automatic and intelligent detection of video's temporal quality which is transmitted by VLC.(3)Research on the technology of video spatial detection based on Siamese Network.In this part the Siamese Network is built to extract the color,texture,contrast and other features of the reference images and the images under test.And then the features for the full reference image quality assessment are concatenated to optimize the network structure fitting images with any size.To improve the applicability of spatial quality detection for videos transmitted by VLC,an alternative method about no reference image quality assessment based on the pseudo reference image is proposed.(4)Research on the technology of video quality assessment based on multi-task deep learning.In this part the video quality assessment network based on spatiotemporal visual sensitivity is analyzed and built to extract spatiotemporal joint features of the video and regress video quality assessment scores.According to the temporal detection and spatial detection model,a feature sharing structure for multi-task learning is designed to improve the effects of feature learning for different tasks.(5)Executing experiments of temporal detection,spatial detection and comprehensive quality assessment of videos transmitted by VLC.In this part the hardware integrated platform and software integrated platform of VLC video transmission quality assessment experiment are introduced,and the application effects of temporal detection,spatial detection and video quality assessment are tested and analyzed comprehensively. |
