Fruit image classification based on improved Res2Net and transfer learning

Aiming at the shortcomings of the traditional fruit image classification algorithm with weak feature learning ability and weak representation of fine-grained feature information, this paper proposes a fruit image classification algorithm based on improved Res2Net with migration learning. First, for the network structure, a dynamic multi-scale fusion attention module is introduced into the residual unit of Res2Net to dynamically generate convolution kernels for images of various sizes, optimize the ReLU activation function by using the meta-ACON activation function, and dynamically learn the linearity and nonlinearity of the activation function to adaptively choose whether to activate the neurons or not; second, a training method based on model migration is used to further improve the efficiency and robustness of classification. The experimental results show that the algorithm proposed in this paper improves the test accuracy on Fruit-Dataset and Fruits-360 dataset by 1.2% and 1% compared with Res2Net, and the recall rate improves by 1.13% and 0.89% compared with Res2Net, which effectively improves the performance of fruit image classification. © 2025 University of Electronic Science and Technology of China. All rights reserved.