HCA-Mamba: a hierarchical cross-attention framework combining Vision Mamba and CNN for night-time image quality...

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Night-time image quality assessment (NTIQA) is challenging due to the need to model both global illumination and local artifacts like noise and glare.
• Current Convolutional Neural Network (CNN) methods excel at local feature extraction but struggle with long-range dependencies.
• Transformer-based methods capture global context but often incur high computational costs.

Purpose of the Study:

• To develop a hybrid framework for NTIQA that effectively integrates global context modeling and local distortion perception.
• To improve the accuracy and efficiency of NTIQA by combining the strengths of Vision Mamba (ViM) and CNN architectures.
• To enhance the model's ability to generalize across different datasets and varying levels of night-time image degradation.

Main Methods:

• Proposed HCA-Mamba framework utilizing a Vision Mamba (ViM) backbone for global context and a parallel CNN branch for local distortion extraction.
• Developed a Multi-Scale Distortion Extractor (MSDE) to compact local degradation cues into distortion-aware tokens.
• Implemented a Local Distortion Injection Module (LDIM) for gated cross-attention, enabling stable interaction between global and local representations within the ViM layers.

Main Results:

• The HCA-Mamba framework demonstrated superior performance on the NNID and EHNQ benchmarks for NTIQA.
• Effective cross-dataset generalization was observed on the NPHD dataset in both intra-dataset and cross-dataset settings.
• Ablation studies confirmed the significant contribution of MSDE and LDIM in improving sensitivity to night-time image distortions.

Conclusions:

• HCA-Mamba offers an effective hybrid approach for NTIQA, successfully balancing global and local feature modeling.
• The proposed method shows strong generalization capabilities, making it suitable for diverse real-world night-time imaging scenarios.
• The integration of ViM and CNN, facilitated by MSDE and LDIM, provides a robust solution for analyzing low-light image quality.