We present a deep network-based blind image quality assessment (BIQA) using two-side pseudo reference (TSPR) images. Different from a traditional reference with the perfect quality, two opposite images are introduced as pseudo references, which are generated via a restoration model and a degradation model, respectively. The bilateral distance (error) maps between the TSPR images and the pristine distorted images are first extracted, and then the regression neural network is utilized to simulate how the human brain processes information, i.e., dialectical thinking, in which both positive information and negative information are taken into consideration for predicting the quality of images. To avoid the complexity and differentiation of each optimization objective in the end-to-end strategy, the step-by-step strategy is adopted to make the model clearer and more explainable. Finally, four common distortions on the LIVE-2D, TID2013, and CSIQ datasets are evaluated using the proposed model, and the experiment results demonstrate the effectiveness and robustness of the algorithm, which delivers superior performance over the state-of-the-art NR methods.

我们提出了一种使用双面伪参考（TSPR）图像的基于深度网络的盲图像质量评估（BIQA）。与具有完美质量的传统参考不同，引入了两个相反的图像作为伪参考，它们分别通过恢复模型和退化模型生成。首先提取TSPR图像和原始失真图像之间的双边距离（错误）图，然后使用回归神经网络来模拟人脑如何处理信息，即辩证思维，其中正面信息和负面信息考虑到预测图像质量。为了避免端到端策略中每个优化目标的复杂性和差异性，采用分步策略使模型更清晰，更易解释。最后，使用提出的模型评估了LIVE-2D，TID2013和CSIQ数据集上的四种常见失真，并且实验结果证明了该算法的有效性和鲁棒性，与最新的NR相比，它具有出色的性能。方法。