Structure-Preserved Self-Attention for Fusion Image Information in Multiple Color Spaces.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning model performance is influenced by color space selection.
• Current methods for integrating color spaces lead to larger models and inefficient information use.

Purpose of the Study:

• To propose an efficient model for fusing image information from diverse color spaces.
• To enhance deep learning recognition performance by optimizing color space utilization.

Main Methods:

• Introduced the structure-preserved self-attention network (SPSANet) with a novel structure-preserved self-attention (SPSA) module.
• Employed a single-head pixel-wise attention mechanism instead of multihead self-attention (MHSA).
• Utilized feature maps from all color space paths for similarity matching and incorporated channel shuffle operations.

Main Results:

• SPSANet demonstrated superior recognition performance across eight common color spaces (RGB, Luv, XYZ, Lab, HSV, YCrCb, YUV, HLS).
• The model achieved enhanced performance with significantly reduced parameters and computational cost.
• The proposed SPSA module effectively mitigates color space dependency and leverages multi-color space advantages.

Conclusions:

• SPSANet offers an efficient and effective solution for multi-color space image information fusion in deep learning.
• The model's design optimizes information utilization, leading to improved recognition accuracy and efficiency.
• This approach provides a promising direction for developing more robust and computationally efficient deep learning models for image recognition tasks.