Distribution Learning Based on Evolutionary Algorithm-Assisted Deep Neural Networks for Imbalanced Image...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Imbalanced image classification presents challenges in generating high-quality and diverse synthetic minority samples.
• Existing methods struggle to balance sample quality with diversity, hindering performance on imbalanced datasets.

Purpose of the Study:

• To propose an evolutionary algorithm-assisted deep distribution learning framework, MEDA_LUDE, for optimizing latent feature distributions.
• To enhance the quality-diversity tradeoff in synthetic minority sample generation for imbalanced image classification.

Main Methods:

• Developed the improved estimation distribution algorithm-based latent feature distribution evolution (MEDA_LUDE) framework.
• Employed a multivariate Gaussian mixture (GM) assumption and a novel four-phase training strategy.
• Introduced a large-margin GM (L-GM) loss for dynamic covariance modeling and a similarity-guided fitness function for MEDA.

Main Results:

• Achieved 95.9% accuracy on MNIST (IR:100), outperforming state-of-the-art by 1.26% on CIFAR-10.
• Improved accuracy by 1.45% on DHU-FD and 0.92% on ALIYUN-FD for industrial fabric defect datasets.
• Demonstrated significant gains in precision (2.5%) and G-mean (1.17%) on DHU-FD, with superior quality-diversity tradeoffs in generated samples.

Conclusions:

• MEDA_LUDE effectively addresses imbalanced image classification by optimizing latent feature distributions.
• The framework demonstrates superior performance on benchmark and real-world datasets, particularly in fabric defect detection.
• The proposed method offers a practical solution for generating high-quality, diverse synthetic minority samples in imbalanced learning scenarios.