Deep Neural Networks for No-reference Image Quality Assessment

• Main Authors: CHENG, FENG-SHIH, 鄭豐時
• Other Authors: KUO, TIEN-YING
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/5zzzxh

碩士 === 國立臺北科技大學 === 電機工程系 === 107 === Image quality assessment (IQA) can automatically evaluate objective image quality scores, and thus can replace the time-consuming subjective quality assessment operations. Benefit from the rapid development and the success of convolutional neural networks, many CNN-based image quality assessment methods have been proposed. Studies have confirmed that adding saliency into the objective IQA algorithm can improve the accuracy of the evaluation as it is used as a visual attention mechanism. Currently, most of the saliency-based IQA models are full-reference methods, which need the information of the undistorted image. In the saliency-based no-reference (NR) IQA method, most of them are evaluated for the specific distortions that cannot be widely used for various types of distortion. The most important thing is that the majority of these methods are patch-wise prediction, which is less consistent with the human visual system that observes the distortion of the entire image and then evaluates the score. We propose an NR IQA method based on deep neural networks to evaluate the quality of the entire image. It consists of three sub-networks. 1) Saliency detection network. It first finds out the area that human eyes will notice first when viewing the image. If the distortion range is in large area, it will also help the detection. 2) The distortion-distribution generating network. It can detect the position and degree of the distortion in images, and can apply to various distortion types with the help of the neural network. Finally, the information obtained from the above two networks is used as a visual aid to 3) the score prediction network to evaluate the quality score of images. The experimental results show that the method in this paper is close to the real score for most of the distorted images in the LIVE database, but it performs poorly on Gaussian white noise. In the TID2013 database, the proposed method has limited ability to predict the scores of various distortions. In the future, it is necessary to consider combining the feature maps in different ways, then input the network, and make different architectural adjustments to the distortion distribution generation network and the score prediction network.