Abstract
Aiming at the issues of high computational cost and limited generalization ability of ResNet50 in classifying images, this study advances an optimization strategy based on transfer learning. The model is initialized with transfer learning to reduce computational burden, and data augmentation techniques are employed to enhance generalization ability. Additionally, label smoothing is introduced to optimize the cross-entropy loss, thereby reducing sensitivity to noisy labels. The training process is further optimized using cosine annealing learning rate decay. Experimental findings reveal that the optimized ResNet50 model achieves a 6.25% improvement in classification accuracy on the CIFAR-10 dataset, validating the validity of the suggested methods.