Research On No-reference Surveillance Image Quality Assessment

With the development of multi-media technology,digital image has been more and more popular in our everyday life and become the important access to information as it is easy to be transported and saved.As a result,Video Surveillance System has been widely used.Many Distortion Types are introduced in the process of collection,compression,transportation and reappearance.The quality of terminal surveillance image will directly affect visual information via eyes even security.Therefore,it is significant to evaluate the quality of image.The most effective way to evaluate is to judge by experts,that is,subjective quality evaluation.However,there are shortcomings like huge human consumption and inefficiency.Consequently,the method is unable to meet the need of massive image evaluation.For example,the quality of image is closely related to the availability of information in widely-used Video Surveillance System.To guarantee the efficacy of the system,it is necessary to manage the quality of image.Since the sharp increase of Video Surveillance System makes traditional method far from the aims,it has become the important direction to use objective method based on computer analysis.The methods are as follows,full reference algorithm which needs origin image as full-reference,reduced-reference algorithm which needs part information of origin image and no-reference algorithm which needs no origin image.In practical application,it is hard to achieve origin distortion image especially surveillance image which has been multiprocessing.Thence no-reference algorithm is the inevitable choice.This article focuses on no-reference algorithm,first studying classic algorithm,then analyzing and comparing their function.Based on existing methods,the article aims at particularity of evaluating surveillance image,introducing 2-D entropy,block analysis and improving the drawbacks of other methods.The hidden features of image are that subjective evaluation is affected by object and area of interest,image is affected by multiple distortion and different distortion types and distribution have different features.This article introduces Probabilistic Latent Semantic Analysis to model the hidden features of image.To obtain potential topics distribution,the article builds and solve the relationship of image document,potential topics and characteristics word.Potential topics can well reflect the degree of influence of noise on image.Via machine learning,the article studies the relationship of potential topics and quality of image and finally can evaluate the quality of surveillance image precisely.