Font Size: a A A

Research On Videos Quality Classification Algorithm Based On Deep Learning

Posted on:2021-08-09Degree:MasterType:Thesis
Country:ChinaCandidate:Q WangFull Text:PDF
GTID:2518306050472584Subject:Master of Applied Statistics
Abstract/Summary:PDF Full Text Request
In recent years,with the rapid development of the security industry,its core video surveillance technology has been widely used in various industries.In addition to some special areas such as transportation,public security,finance,banking and so on.Monitoring systems have been installed in many shopping malls,office buildings,hotels,factories,communities,and even homes.However,due to various factors,video images often inevitably suffer from distortion problems such as signal loss,picture freeze,abnormal brightness,abnormal color,and excessive noise,which affects the user's subjective feelings and access to information.Therefore,it is an important and urgent problem to realize accurate image quality diagnosis.The location of video surveillance is ever-changing,and the traditional algorithms with poor generalization performance can no longer perfectly solve such video quality classification problems.In recent years,deep learning algorithms have developed rapidly,and their powerful expression and learning capabilities can handle different complex environments.Therefore,this paper uses Caffe as an experimental platform to propose a quality classification algorithm based on deep learning for its video information.The main work contents are as follows:First of all,the basic structure and training foundation of the convolutional neural network are studied in detail,and the common improvement directions of the convolutional neural network model are proposed.Secondly,for no signal,color cast,snowflake,occlusion,stripes,jitter,blur,and bright 10 types of abnormal attributes,dark,and black screen were defined.A total of 22,503 images were collected to create image tags,a video quality image library was established,and the data was processed accordingly.Thirdly,the traditional training method of deep learning was improved.The warm up method was used to study the classification effect of different convolutional neural networks on the video quality diagnostic data set.A comparison found that the improved training method not only increased the accuracy of the resnet18 model by 17% And the recognition time is reduced by 1ms,so Res Net18 is initially selected as the basic model.Then,by improving the loss function and network structure,an improved Res Net18 model is proposed,and on the video quality diagnostic image library,the improved model and the basic Compared with the ResNet18 network,the results show that the algorithm proposed in this paper not only shortens the model training time to one-fifth of the original,but also greatly reduces the amount of network parameters,and the classification accuracy is further improved by about 4%;finally,in the image The classified public data set cifar10 verifies the video quality classification algorithm proposed in this paper.The experimental results show that the proposed algorithm does have better recognition effect.
Keywords/Search Tags:video image classification, deep learning, Residual Network, multi label classification, Caffe framework
PDF Full Text Request
Related items