Learning from Strong to Weak an Enhanced Quality Comparison Network via Efficient Transfer Learning

Abstract

Blind image quality assessment (BIQA) faces challenges with high-resolution image processing and the impracticality of deploying large, computationally demanding models, while lighter models suffer from significant performance gaps. To address these issues, we propose a knowledge transfer strategy for heterogeneous models. Specifically, we introduce a large-parameter teacher model, Enhanced IQANet (EIQANet), which integrates rich image features, utilizes a composite loss function to optimize BIQA performance, and implements a dynamic patch cropping scheme tailored for high-resolution image processing. These combined enhancements lead to superior performance on the UHD-IQA dataset. Simultaneously, we present the Enhanced Quality Comparison Network (EQCNet), a lightweight and deployable student model featuring an efficient backbone that reduces computational complexity and an upsampling strategy designed to minimize artifacts during image scaling, thereby enhancing ranking accuracy in quality assessment. To bridge the performance gap between EIQANet and EQCNet, we employ a multi-stage knowledge transfer strategy, including pre-training, fine-tuning, and calibration. This approach ensures effective knowledge transfer between heterogeneous models and optimizes feature alignment, promoting a well-structured embedding space. Our method achieves outstanding results on the UHD-IQA validation benchmark.