Image Quality Assessment Method Based On Machine Learning

With the advent of the era of big data,more and more digital images are being shared on the Internet.Therefore,how to effectively process and transmit images,determine the quality of image compression algorithms,and accurately evaluate the image quality have become urgent issues to be studied.Traditional weighted methods require artificial weighting and cannot reflect the importance of each feature Therefore,more and more machine learning methods are utilized in image quality evaluation.Through the study of human visual system,machine learning and deep learning,we proposed two methods for evaluating image quality.A FR-IQA method is proposed,which is based on the masking texture feature.Through the study of the characteristics of human visual system,the masking effect is an important characteristic of HVS,and it is mainly determined by the complex texture and background brightness.We proposed a texture feature extractor method to mimic the masking effect of human visual system.Combing with gradient features and color features,we use random forest technique to build the regression model and predict image quality scores.The overall performance experimental results on the five image quality assessment databases show that the four evaluation indexes of the proposed TCQI is better than the eight mainstream FR-IQA methods.And the Pearson correlation coefficient of TCQI on the TID2013 database reached 0.9448.Cross-database validation and complexity experiments show that the proposed method has good robustness and low computational complexity,and it can be used in real-time digital image quality detection system.A blind image quality method(CNN-CroW)is proposed,which is based on convolutional neural network.Because traditional convolutional neural network has strict limitations on the size of input image,we firstly split the input image signals into multiple non-overlapping blocks and the convolution features of each blocks are extracted as local features.Then the convolution features are aggregated using Cross-dimensional weighting method to obtain the global features.It can avoid the size limit of input image.Then the Crow features are connected to the full-connected layer for modeling and predicting the image quality score finally.Compared to the traditional learning-based methods which need to extract feature artificially,the proposed CNN-CroW model inputs the original color image directly and there is no limit in the size of input image.The experimental results show that the proposed CNN-CroW method can predict the image quality scores accurately on each databases and have a high consistency with subjective evaluation.