Conv-Former: A Novel Network Combining Convolution and Self-Attention for Image Quality Assessment

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• No-reference image quality assessment (NR-IQA) is crucial for evaluating image authenticity and distortion.
• Existing NR-IQA methods struggle with both authentic and synthetic image distortions.
• Accurate image quality assessment (IQA) models require robust perceptual mechanisms.

Purpose of the Study:

• To propose a novel network, Conv-Former, for NR-IQA.
• To enhance representation learning for improved image content understanding.
• To achieve state-of-the-art performance on both authentic and synthetic image databases.

Main Methods:

• Utilizing a multi-stage transformer architecture inspired by ResNet-50 for perceptual mechanisms.
• Implementing adaptive learnable position embedding for arbitrary image resolutions.
• Introducing a new transformer block (TB) combining long-range dependencies and local information perception (LIP).
• Employing dual path pooling (DPP) to preserve contextual image quality information during feature downsampling.

Main Results:

• Conv-Former outperforms state-of-the-art methods on authentic image databases.
• Conv-Former achieves competitive performance on synthetic image databases.
• Experimental results demonstrate strong fitting performance and generalization capability.

Conclusions:

• Conv-Former offers a robust solution for NR-IQA across diverse image distortion types.
• The proposed network effectively integrates convolutional and self-attention mechanisms for enhanced IQA.
• Conv-Former shows significant potential for real-world image quality evaluation applications.