Compared with general image classification, fine-grained image classification pays more attention to the local features of objects, but due to the small differences in the classified objects, the distinction between some categories is low, which increases the difficulty of classification. This paper aims to optimize fine-grained image classification using the DenseNet architecture to improve the accuracy and robustness of the model. First, a fine-grained image dataset containing a variety of species is collected, and data augmentation is performed to expand the training samples, using techniques such as random cropping, rotation, and color transformation. Next, DenseNet is used as the base model, and the pre-trained weights are loaded through transfer learning, and further fine-tuned on a specific dataset. At the same time, regularization techniques such as Dropout and Batch Normalization are introduced to reduce overfitting. The experimental results show that the classification accuracy of the optimized DenseNet model on the test set reaches 91.3%, and the F1-score is 0.89, indicating that the model performs well in fine-grained classification tasks. The effective combination of DenseNet’ s deep feature extraction capabilities and optimization strategies has significantly improved the performance of fine-grained image classification and provided strong support for applications in related fields.