The Research Of Key Problems In Image Quality Assessment Based On PLSA

Nowadays with image information technology widely applied into many fields, the requirement of no-reference image quality assessment increases with each passing day. Comparing with the development of reference image quality assessment, the research of no-reference image quality assessment is still in the initial stage. However, the introduce of no-reference image quality assessment establishes a new way and makes good progress. But at the same time, there is still some key problems to be solved, such as the uncertainty of evaluation in PLSA, too simple the semantic similarity measurement methods are and the access to underlying multiple distortion feature of distorted images. Thus solving these problems is the key to increase the performace of no-reference image quality assessment.In this article, I solve the above three problems respectively by K-fold-cross-validation, KL distance, Hellinger distance and the algorithm of multidimensional feature extraction of distorted images based on the image quality assessment with the PLSA algorithm, and obtain the results.Afterwards, I make a comparison with the result without using K-fold-cross-validation, the result when semantic similarity is cosine similarity and the result of three-dimensional feature such as blur, compression and noise. Then the conclusions are obtained.Firstly, from the experiment, it can be seen that K-fold-cross-validation can expand the number of samples in a certain extent, obtain more stable model and play an important role in the training process. Secondly, from the comparison of the results of KL distance, Hellinger distance and cosine similarity, it can be seen that KL distance is not suitable for no-reference image quality assessment, but Hellinger distance and cosine similarity can be both applied into no-reference image quality assessment as the semantic similarity.At last, as for multidimensional feature extraction of distorted image, not only the result of that is more obvious and suitable than that of three-dimensional feature extraction in the application of no-reference image quality assessment, but also the algorithm measures images more widely.