SF-IQA: Quality and Similarity Integration for AI Generated Image Quality Assessment

Zihao Yu, Fengbin Guan, Yiting Lu, Xin Li, Zhibo Chen; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 6692-6701

In recent years the rapid development of Artificial Intelligence (AI) has facilitated the widespread use of AI-Generated Images (AIGIs) a subset of Artificial Intelligence Generated Content (AIGC). However there are prevalent issues associated with AIGIs notably the unsatisfied quality of the generated images and the misalignment between the generated images and their corresponding textual prompts. These challenges underscore the importance of Image Quality Assessment (IQA) in the field of AIGIs to provide more precise quality predictions that are consistent with human perception. Responding to this need we introduce SF-IQA a novel AIGC image quality metric that integrates quality and similarity in a score fusion manner. Specifically we employ a multi-layer feature extractor and fusion module to extract and aggregate the local and global-level features facilitating the excavation of quality-aware features. For image-text similarity we fine-tuned a strong vison-language model based on a powerful perceptual-aware image-text alignment prior. With the assistance of score fusion manner our proposed SF-IQA obtains state-of-the-art results on AGIQA-3K benchmarks and achieves 4th place in the NTIRE 2024 Quality Assessment of AI-Generated Content Challenge.