Cross-and-Diagonal Networks: An Indirect Self-Attention Mechanism for Image Classification

Area of Science:

• Computer Vision
• Deep Learning
• Image Classification

Background:

• Fully convolutional neural networks (FCNs) excel at local image features but struggle with global context.
• Self-attention mechanisms capture global context but are computationally intensive.
• Existing methods face limitations in balancing local detail preservation with global context awareness and computational efficiency.

Purpose of the Study:

• To introduce Cross-and-Diagonal Networks (CDNet), an innovative architecture for efficient image classification.
• To address the limitations of FCNs and self-attention mechanisms in capturing global image context.
• To enhance computational efficiency while maintaining or improving classification performance.

Main Methods:

• Proposing Cross-and-Diagonal Networks (CDNet) with a novel "cross and diagonal attention" mechanism.
• Implementing an indirect self-attention approach to establish long-range pixel dependencies.
• Integrating the attention mechanism as a module into Convolutional Neural Networks (CNNs).

Main Results:

• CDNet effectively captures global image information while preserving local details.
• The indirect self-attention mechanism significantly enhances network capacity and computational efficiency.
• State-of-the-art performance achieved on three benchmark image classification datasets.

Conclusions:

• CDNet provides an efficient and effective solution for image classification by capturing global context.
• The proposed indirect self-attention mechanism offers a computationally lighter alternative to traditional methods.
• CDNet demonstrates superior performance and efficiency, making it a valuable advancement in computer vision.