ملخص: Despite remarkable success has been achieved by convolutional neural networks (CNNs) in no-reference image quality assessment (NR-IQA), there still exist many challenges in improving the performance of IQA for authentically distorted images. An important factor is that the insufficient annotated data limits the training of high-capacity CNNs to accommodate diverse distortions, complicated semantic structures and high-variance quality scores of these images. To address this problem, this paper proposes a hierarchical curriculum learning (HCL) framework for NR-IQA. The main idea of the proposed framework is to leverage the external data to learn the prior knowledge about IQA widely and progressively. Specifically, as a closely-related task with NR-IQA, image restoration is used as the first curriculum to learn the image quality related knowledge (i.e., semantic and distortion information) on massive distorted-reference image pairs. Then multiple lightweight subnetworks are designed to learn human scoring rules on multiple available synthetic IQA datasets independently, and a cross-dataset quality assessment correlation (CQAC) module is proposed to fully explore the similarities and diversities of different scoring rules. Finally, the whole model is fine-tuned on the target authentic IQA dataset to fuse the learned knowledge and adapt to the target data distribution. Experimental results show that our model achieves state-of-the-art performance on multiple standard authentic IQA datasets. Moreover, the generalization of our model is fully validated by the cross-dataset evaluation and the gMAD competition. In addition, extensive analyses prove that the proposed HCL framework is effective in improving the performance of our model.

لغة: إنجليزية
الموضوع الإعلام الآلي

الكلمات الدالة:
Prior knowledge
No-reference image quality assessment
Hierarchical curriculum learning
Cross-dataset quality assessment correlation