No-reference image quality assessment based on deep learning method

In this work, we adopt the use of deep learning method for no-reference image quality assessment. With the development of deep neural networks technology, foundational and deep features of images could be captured without much prior knowledge. So a sparse autoencoder (SAE) was trained to express a 32 × 32 pixels image into a feature vector. Then the original images were cut into serial sub-images with the size of 32 × 32 pixels which can fix the input size of SAE. After that, the features vector of each sub-image was extracted separately and the information was fused with two strategies for the image quality assessment task. The best strategy in this work is that each sub-score is calculated by a Support Vector Regression (SVR) machine with the input of sub-image feature vector and estimate the image quality by averaging the scores to get the final score for the original image. Moreover, the effectiveness of our proposed method was confirmed by the experimental results in the TID2013 image quality assessment database.