BiRQA: Bidirectional Robust Quality Assessment for Images

A research team from an unspecified institution has introduced BiRQA (Bidirectional Robust Quality Assessment), a novel model designed to solve critical weaknesses in Full-Reference Image Quality Assessment (FR IQA). Current neural metrics for comparing a processed image to its original are notoriously slow and vulnerable to manipulation, limiting their use in real-time systems like video compression or reliable evaluation of AI-generated imagery. BiRQA addresses this by combining a compact, efficient architecture with a groundbreaking defense strategy, positioning itself as the only FR IQA model to offer competitive accuracy, real-time speed, and strong adversarial resilience simultaneously.

The model's technical innovation is twofold. First, its bidirectional multiscale pyramid architecture uses a bottom-up attention module and a top-down cross-gating block to efficiently blend fine-scale details with semantic context. Second, and more crucially, the team developed 'Anchored Adversarial Training,' a theoretically grounded method that uses clean 'anchor' samples and a ranking loss to bound prediction errors during attacks. The results are compelling: BiRQA matches or outperforms previous state-of-the-art models on five public benchmarks while processing images approximately three times faster. Most impressively, under unseen white-box adversarial attacks on the KADID-10k dataset, it increased the key robustness metric (SROCC) from a range of 0.30-0.57 to 0.60-0.84—a massive leap in reliability. This combination of speed and security makes BiRQA a practical tool for developers working on robust image compression, restoration pipelines, and trustworthy evaluation of generative AI models like Stable Diffusion or DALL-E.