Abstract
A highly efficient deep fully convolutional neural network (DFCN) for image quality assessment (IQA) is designed in this paper. The DFCN consists of two branches, one scoring local patches and the other estimating the weights of local patches to enhance quality prediction. Then, the DFCN outputs quality score of the whole image with aggregate weighted average pooling. There are no fully connected layers in the DFCN, resulting in far fewer parameters. In addition, the network model utilizes multiscale images as inputs to enrich the extracted distortion information. Furthermore, the parameters of the model are optimized in two steps to reduce the requirement for computing power and the risk of overfitting. The parameters of the shared layers and the quality module are optimized firstly, and then, the parameters of the weight module are optimized with the designed loss function. The extensive experimental results show that the proposed DFCN outperforms other competing IQA methods and has strong generalization ability.
Highlights
Advances in Multimedia learning has been widely applied to image quality assessment (IQA) metrics
Motivated by Kang and Wang [2] and Ma et al [3], we proposed a simple and highly efficient deep fully convolutional neural network model; one branch predicts primary scores of local patches, and the other branch enhances the predicted quality by estimating the weights for the local patches
Four scale images are generated with the image pyramid. e primary quality scores and the weights for the local patches are obtained through trained modules g1 and g2, respectively. en, the more accurate quality scores of the local patches are computed via the Hadamard product
Summary
We supposed that the local patches are evenly assigned quality labels from the whole annotated images. Equation (6) describes the average error between the predicted quality score and ground truth score. To mitigate overfitting of the model, the regularization constraint item is introduced as. Λ is a balance parameter between two items. E assumption that local quality is uniformly assigned over the distorted image causes a tremendous amount of label noise. We design the other branch, weight module, which learns the weights for the local patches to enhance quality prediction. The parameters of module g2 are optimized with a unified loss function after w0 and w1 are fixed, which is described as
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
