Grapevine Leaves Recognition Based on IP-ShuffleNet

How to identify grape varieties based on the shape of the leaves is of great significance for grape breeding and evolution. This article takes the leaves of different grape varieties as the object, addressing these issues of high model complexity and poor generalization ability common in current deep learning image classification tasks. To reduce model complexity, a lightweight neural network model, IP-ShuffleNet, based on an improved ShuffleNetV2, is proposed. This enhanced model employs a multi-channel feature fusion module to replace pooling layer, introduces the Squeeze-and-Excitation (SE) and A Simple, Parameter-Free Attention Module (SimAM) attention mechanisms, and the Context Broadcasting (CB) module to assist the model feature learning. Meanwhile, the Deep Convolutional Generative Adversarial Neural Network (DCGAN) is improved for data augmentation on a publicly available five-class Grapevine Leaves Image, enhancing the model’s capacity, generalization ability, and spatial information interaction. Experimental results show that the proposed model achieves an accuracy of 92.37% when directly trained on the augmented dataset and reaches a peak accuracy of 98.1% through transfer learning with pre-trained weights, which provides a more effective solution for grape variety.