DH-GarlicNet: a precise identification method for garlic damage based on the improved residual network.

With the increasing level of agricultural automation, the detection of damaged garlic has become an important task to ensure food quality and improve production efficiency. This paper presents a novel damaged garlic detection model -DH-GarlicNet. This model is based on the ResNet34 architecture and innovatively integrates depthwise convolution, SE attention module, and SiLU activation function, aiming to significantly improve the accuracy of damage detection. With the introduction of depthwise convolution, the model can efficiently extract subtle features in garlic images; the SE module enhances the focus on critical regions; and the SiLU activation function optimizes the representation of nonlinear features. The experimental results show that the DH-GarlicNet model achieves 98.95% detection accuracy and 0.0793 loss in the damaged garlic detection task, which improves the accuracy by 7.37% and reduces the loss by 0.2324 with respect to the original model, and significantly outperforms the traditional detection algorithm. The Grad-CAM visual analysis demonstrates the model's ability to focus on the damaged regions, proving its advantages in target recognition and local feature learning. This study provides an accurate solution for the automated detection of damaged garlic, which has a strong application prospect.