Image Quality Assessment Using Combination of Deep Convolutional Neural Networks

Easy access of hand-held image capturing devices has increased storage and transmission of visual digital data immensely. Hence, device assisted evaluation of quality of images without any prior information is of great interest. Historically, handcrafted natural scene/image statistics have been widely used for image quality assessment (IQA). Recent advances have shown that deep convolutional neural networks (CNN) extract fundamental information about images which can be utilized for IQA. Distortion information, responsible for quality degradation, is an important aspect for IQA and affects the quality level as well. In the proposed model, a combined deep CNN architecture is trained for distortion identification and quality prediction. The proposed combination of deep CNNs with shared initial layers is fine tuned for both the tasks, and extracted features are combined to predict the image quality scores via kernel ridge regression. Performance evaluation is reported on LIVE image quality assessment database. Proposed approach not only performs statistically better than the full reference IQA measures structural similarity index and feature similarity index, but also shows improvement over the CNNs fined tuned only for quality prediction.