DMFGAN: a multifeature data augmentation method for grape leaf disease identification.

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Deep learning for grape leaf disease identification requires large datasets.
• Large datasets increase computational costs and risk pattern collapse during training.

Purpose of the Study:

• To propose a Depth-separable Multi-feature Generative Adversarial Network (DMFGAN) for enhancing grape leaf disease data.
• To address limitations of existing generative models in terms of image quality, feature learning, and training stability.

Main Methods:

• Developed a Multi-feature Extraction Block (MFEB) using four-channel feature fusion.
• Designed a depth-based D-discriminator to enhance discriminative power and reduce parameters.
• Implemented SeLU activation function and an MFLoss function with a gradient penalty to mitigate mode collapse.

Main Results:

• DMFGAN generated higher quality and more diverse grape leaf disease images compared to other generative adversarial networks.
• The proposed method reduced model parameters and mode breakdown occurrences during training.
• Validation using a recognition network confirmed the effectiveness of the enhanced data.

Conclusions:

• DMFGAN offers an effective solution for enhancing grape leaf disease datasets.
• The model achieves superior performance with reduced computational resources and improved training stability.
• This approach holds significant potential for practical applications in disease identification.