D2S-DiffGAN: a novel image classification model under limited labeled samples.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Data scarcity is a major limitation for deep learning in image classification.
• Existing Generative Adversarial Network (GAN) models primarily focus on spatial domain features, neglecting frequency domain information.
• Current models often fail to differentiate the contribution of real and generated samples during training.

Purpose of the Study:

• To propose a fully supervised image classification model (D2S-DiffGAN) effective under limited labeled samples.
• To enhance data generation by incorporating both spatial and frequency domain constraints.
• To improve model training by utilizing a differentiated loss function.

Main Methods:

• A dual-domain synchronous GAN (DDSGAN) was developed to generate diverse and realistic samples by constraining both spatial and frequency domains.
• A multi-branch feature extraction network (MBFE) with an attention module was designed to capture and fuse multi-dimensional features.
• A differentiated loss function (DIFF) was proposed to assign distinct weights to real and generated samples.

Main Results:

• The D2S-DiffGAN model achieved good classification accuracy on the SVHN and CIFAR-10 datasets despite limited labeled samples.
• The DDSGAN effectively generated samples with both visual realism and consistent frequency domain energy distribution.
• The MBFE and DIFF components contributed to enhanced feature representation and optimized model training.

Conclusions:

• The proposed D2S-DiffGAN model demonstrates significant effectiveness in image classification tasks with limited labeled data.
• Integrating dual-domain generation and differentiated loss functions offers a promising approach to overcome data scarcity challenges in deep learning.
• The model's ability to leverage both spatial and frequency domain information enhances the quality of generated data and classification performance.