Color Image Quality Assessment with Multi Deep Convolutional Networks

Abstract

Listen

Translate article

The existing objective natural Image Quality Assessment (IQA) methods are mostly designed for grayscale images. These methods ignore the correlations between color information and quality distortions, and lack of investigating the feature representations of color information. Consequently, the prediction accuracy and generalization performance on color images are not satisfactory. In view of this, a multi deep Convolutional Neural Network (CNN) architecture is designed for color image quality assessment, which deeply explores the color information in different color spaces. Firstly, a group set of component maps are generated with multi-scale transformation and multi-color-space transformation on the color image. Secondly, the general CNN structures are adopted and improved to learn effective feature representations related to image quality. Each component map is fed to a single CNN, thus forming a multi-CNNs model. Thirdly, each CNN is trained using transfer learning method with pretrained models on large scale datasets. Finally, multiple output feature eigenvectors are fused and mapped to the subjective quality scores to construct the quality prediction model. The systematic experimental results on multiple artificial databases and the real-world database showed that the proposed prediction model outperforms existing IQA methods.