No-Reference Image Quality Assessment Based On Deep Features Of Enhancement

With the rapid development of internet technology and transmission equipment,Image data is widely used in industrial production,public security and other fields.How to express the image quality in the process of capturing,transmitting and storing images has been an important research content.Image quality assessment plays an important role in image processing and performance evaluation.This paper studies the effectiveness of attention feature enhancement in image quality evaluation and establishes the relationship between objective algorithm and subjective perception.Firstly,a no-reference image quality assessment algorithm with multi-level features enhancement(MFEQA)is proposed.In this paper,the feature extraction network guided by the proposed feature enhancement module is used to extract the features of the distorted images,and the extracted features are strengthened in channel and position dimensions to obtain the region of interest.Then the feature fusion is carried out through the top-down feature fusion method to enrich the feature information.Finally,the fused features were input to the quality regression model guided by the improved super network to predict the image quality.In MFEQA algorithm,the feature enhancement module is used to enhance the extracted features in channel and position dimensions according to the visual characteristics,and obtain feature diversity information.Using the super network guide quality regression model,we can build the regression model on the basis of understanding the image content to strengthen the perception of the model content.Secondly,an authentically no-reference image quality assessment algorithm based on selfattention Transformer encoder is proposed(SFIQA).The authentically distorted image is characterized by non-uniformity of distortion.In this paper,the local features of distorted image is obtained by feature extraction network.The multi-head self-attention mechanism can establish the relationship between the feature information of different locations.So that the model can capture the long-distance dependence relationship and obtain more comprehensive global features.Then we input the feature into the regression network which consists of full connected layers,and build up the relation between the characteristic and the subjective score to predict the quality of the distorted image.SFIQA algorithm not only obtains local features of distorted image,but also uses multi-head self-attention mechanism to establish relationships between different location features,which strengthens the representation ability of complex image data and describes global information of image in more detail.Effectively improve the accuracy performance of the algorithm model.Finally,we carried out experiments on the proposed MFEQA algorithm and STIQA algorithm in several large public databases.The SROCC of MFEQA algorithm is 0.952 and PLCC of MFEQA algorithm is 0.961 in the CSIQ database.And its performance indicators are better than mainstream no-reference methods.The overall performance is good.The SROCC of SFIQA algorithm in realistic distortion databases LIVEC and Kon IQ-10 k is 0.867 and 0.916,respectively,which is improved compared to other methods.Meanwhile,it has stable performance and high quality perception ability in synthetic distortion databases.The experimental results show that the two algorithms have excellent performance and consistent with subjective perception.