Quality Evaluation of Light Field Sub-aperture Images Based on Deep Learning

Abstract:
Light field images (LFIs) has emerged as a new technology allowing to capture richer visual information from our world. However, the LFIs will meet different degrees of distortion during transmission and imaging display. An effective light field image quality assessment (IQA) index is necessary for the development and application of the light field. We propose a deep learning method based on light field sub-aperture images (SAIs) to solve the problem of quality evaluation of LFIs, called SAIs hyper-IQA Net, it adopts hyper-IQA as backbone network to extract features and the sub-aperture images are used as the input of hyper-IQA according to the image stacks in four different directions. Finally, the feature vectors are combined together for the regression of the final quality score through a fully connected layer. Our proposed model shows that the result has a good consistency with human visual system (HVS) when comparing with existing metrics.