RGSGAN-MACRNet: A More Accurate Recognition Method for Imperfect Corn Kernels Under Sample-Size-Limited Conditions.

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Accurate recognition of imperfect corn kernels is crucial for quality control in agriculture.
• Limited sample sizes severely degrade the performance of current recognition models.

Purpose of the Study:

• To develop an advanced recognition framework to overcome data scarcity challenges in imperfect corn kernel identification.
• To enhance the accuracy and robustness of corn kernel classification under sample-size-limited conditions.

Main Methods:

• Integration of a Residual Generative Spatial-Channel Synergistic Attention Generative Adversarial Network (RGSGAN) for high-quality sample generation and dataset expansion.
• Utilization of a Multi-Scale Asymmetric Convolutional Residual Network (MACRNet) for multi-scale feature fusion and capturing subtle variations.
• Implementation of Wasserstein distance with gradient penalty in RGSGAN for stable and effective generative training.

Main Results:

• The proposed RGSGAN-MACRNet framework achieved a classification accuracy of 98.813%.
• Demonstrated significant improvements over established models like ResNet18, EfficientNet-v2, ConvNeXt-T, and ConvNeXt-v2.
• Outperformed models trained on the original dataset by 5.29%, highlighting its effectiveness with limited data.

Conclusions:

• The RGSGAN-MACRNet framework effectively addresses the challenge of data scarcity in imperfect corn kernel recognition.
• The synergistic combination of generative adversarial networks and multi-scale convolutional networks offers a robust solution for agricultural image analysis.
• This approach significantly alleviates the adverse impact of limited data on recognition accuracy, paving the way for improved agricultural quality control.