Few-shot image classification based on class-irrelevant feature decoupling graph neural network

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Similarity measurement is crucial for image classification but has limitations.
• Existing methods can be misled by irrelevant background features, affecting classification accuracy.
• The coupling of irrelevant background with classification objects leads to attention deviation.

Purpose of the Study:

• To propose a novel framework, the Class-irrelevant Feature Decoupling Graph Neural Network (CFDGNN), to address attention bias in image classification.
• To overcome limitations of similarity measurement methods and class-irrelevant feature coupling.
• To improve the accuracy and attention mechanisms of vision models.

Main Methods:

• The proposed CFDGNN framework integrates three modules: Adaptive Class Salience Channel Weighting (ACS), Main Object Focus Spatial Attention (MFS), and Feature Decoupling Graph Neural Network (FDGNN).
• ACS module prioritizes task-relevant salient features and suppresses irrelevant ones.
• MFS module filters spatial noise and optimizes object texture.
• FDGNN module prevents attention from being misled by simple, irrelevant information.

Main Results:

• Experimental results on miniImageNet, CIFAR-FS, and CUB-200-2011 datasets demonstrate superior classification accuracy compared to existing algorithms.
• The proposed algorithm effectively corrects the attention areas of vision models.
• Visualization experiments confirm the ability to retain main object textures while filtering out irrelevant background textures.

Conclusions:

• The CFDGNN framework offers a significant advancement in computer vision for accurate image classification.
• By decoupling class-irrelevant features, the model achieves more reliable and focused attention.
• This approach enhances the interpretability and robustness of deep learning models in computer vision tasks.