Scholarlink

With the rapid advancement of artificial intelligence, quality assessment of AI-generated content (AIGC) has emerged as a critical research challenge. However, traditional image quality assessment (IQA) methods exhibit limitations in capturing the complex relationship between generated images and their corresponding text prompts, which results in suboptimal performance for AIGC Image Quality Assessment (AIGCIQA).To address these challenges, we propose a dual-branch architecture comprising a Transformer-based text feature extraction branch and a CNN-based image feature extraction branch. Within the shallow convolution neural network (CNN) layers, we incorporate text-guided deformable convolutions to align visual features with text features and enhance visual saliency.The aligned features are then fused, and the final image quality score is obtained by a dedicated module.In addition, we introduce a self-supervised pre-training strategy that leverages intermediate image sequences generated during the Text-to-Image (T2I) process. A full-reference image quality assessment (FR-IQA) model is employed to assign pseudo scores to these images, which serve as a form of supervision to progressively refine during pre-training.Experimental results demonstrate that our method achieves state-of-the-art performance on three mainstream AIGCIQA datasets.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-modal Hybrid Network for AIGC Image Quality Assessment with Self-supervised Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pan Hong,
  • Qingbing Sang

摘要

With the rapid advancement of artificial intelligence, quality assessment of AI-generated content (AIGC) has emerged as a critical research challenge. However, traditional image quality assessment (IQA) methods exhibit limitations in capturing the complex relationship between generated images and their corresponding text prompts, which results in suboptimal performance for AIGC Image Quality Assessment (AIGCIQA).To address these challenges, we propose a dual-branch architecture comprising a Transformer-based text feature extraction branch and a CNN-based image feature extraction branch. Within the shallow convolution neural network (CNN) layers, we incorporate text-guided deformable convolutions to align visual features with text features and enhance visual saliency.The aligned features are then fused, and the final image quality score is obtained by a dedicated module.In addition, we introduce a self-supervised pre-training strategy that leverages intermediate image sequences generated during the Text-to-Image (T2I) process. A full-reference image quality assessment (FR-IQA) model is employed to assign pseudo scores to these images, which serve as a form of supervision to progressively refine during pre-training.Experimental results demonstrate that our method achieves state-of-the-art performance on three mainstream AIGCIQA datasets.