Structural prior-driven feature extraction with gradient-momentum combined optimization for convolutional neural...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Prior information in image classification can improve feature learning but often overlooks feature variability.
• Feature inconsistency leads to decreased accuracy and model instability in image classification tasks.

Purpose of the Study:

• To propose a novel method, structural prior-driven feature extraction with gradient-momentum (SPGM), to enhance image classification accuracy and stability.
• To address limitations of existing methods by focusing on consistent feature learning and precise parameter updates.

Main Methods:

• SPGM utilizes structural prior-driven feature extraction (SPFE) to generate structural information from multi-level features and original images, creating prior knowledge for consistent feature learning.
• An integrated gradient-momentum optimization (GMO) strategy dynamically adjusts parameter updates based on gradient and momentum interactions for precision.

Main Results:

• Experiments on CIFAR10 and CIFAR100 datasets show SPGM significantly reduces the top-1 error rate in image classification.
• The proposed SPGM method demonstrates enhanced classification performance and stability compared to state-of-the-art methods.

Conclusions:

• SPGM effectively improves image classification accuracy and model stability by integrating structural priors and advanced optimization.
• The method offers a promising direction for developing more robust and accurate image classification systems.