An End-to-End Perceptual Quality Assessment Method via Score Distribution Prediction

Image quality assessment (IQA) has become a rapidly growing field of technology as it automatically predicts the perceptual quality, which is of vital importance for consumer-centric services. However, most existing IQA algorithms focus on predicting the mean opinion score regardless of the inevitable opinion diversity. To address this shortcoming, in this paper, we propose to predict the distribution of opinion scores via an end-to-end convolutional neural network. The network is based on a pre-trained ResNet with 50 layers and a novel Statistical Region-of-Interest (ROI) Pooling layer is introduced for lower model complexity, which enables effective training with few datum. Meanwhile, instead of using traditional mean-square-error as loss function, our model is trained with cross-entropy loss, which is more suitable for probability distribution learning. Extensive experiments have been carried out on ESPL-LIVE HDR datasets with highly diverse opinion scores. It is shown that the statistical ROI Pooling is more efficient than traditional ROI Pooling layers and classical dimensionality reduction of principle component analysis. And the proposed algorithm achieves superior performance than state-of-the-art label distribution learning methods in terms of six representative evaluation metrics.