Research On Quality Perception Model And Retrieval Method Based On Visual Information

With the rapid development of Internet and computer vision,big data eral of visual information are flooding in.How to search information quickly and accurately from tremendous data related with visual information has become a very meaningful and challenging research hotspot,and the quality of visual information is the core issue.The quality of visual information directly influences display of information,the accuracy of retrieval,the possibility of knowledge discovery,and then affects the reliability of the artificial intelligence illation.In addition,The increasing amount of visual data on the social networking,such as image,video,indicates that the effective retrieval technology is becoming more and more important.Therefore,this thesis focuses to the statistical properties of visual information,the design of the quality evaluation criteria and retrieval method based on visual information.The main work and innovation points include the following aspects:(1)The existing no-reference video quality assessment based on machine learning methods need to use a lot of subjective scores for training and lead to the problem of high complexity,a blind video quality assessment was proposed.The key of this algorithm is to use Difference of Gaussian filter to extract the video structure feature vector,and to establish multi-level quality perception center collection,which is the video space perceived quality assessment collection.Then,the classification threshold of motion vector was obtained by using clustering algorithm,and the motion perception factors are acquired.Finally,the video objective quality is obtained combining temporal and spatial domain results.Experimental results show that the proposed method is better than other no-reference evaluation algorithms,which is easy to implement.(2)Considering the key influential factors of stereo video Quality of user Experience synthetically,a multi-index 3D QoE evaluation model is proposed,which is for the network transmission.This model first analyzes the four categories of factors.Then,Fuzzy Analytic Hierarchy Process algorithm is used to analyze the influential factors of stereo video QoE hierarchically,and establish evaluation index system.Finally,through calculating the weight of each index,the final evaluation model is obtained.The network simulation experiments and the subjective evaluation scores are employed for verifying the influence of each index on stereo video quality of experience.(3)The noise and incomplete correspondence between the images and the texts give rise to the difficulty of tag-based image retrieval precisely.Therefore,a social image retrieval model based on the hypergraph learning is proposed,which simultaneously utilizes visual feature,textual content and social link information to establish the multi-type relevance between images.Then,an alternating optimization algorithm is used to update the weight of hyperedges.The effect of different edges can be adaptively modulated in the constructed hypergraph.Furthermore,the popularity degree of the image is employed to re-rank the retrieval results.This method is applicable to social media interactive platform(such as Sina weibo).Through using the relationship between the image visual feature,auxiliary text information and users' personal information and interest,the retrieval results meet the user personalization requirements.