模型训练优化策略：AdamW优化器、冻结训练和数据增强

The model was trained using the AdamW[19] optimizer with an initial learning rate of 1e−3, weight decay factor of 0, and a batch size of 30 epochs. The first 15 epochs were frozen training, where the first two layers of the backbone network that did not include cbam were frozen. The next 15 epochs were normal training with a cosine learning rate decay strategy, where the learning rate decreased by 0.9 every 1.3 epochs. Data augmentation techniques such as rotation, cropping, scaling, and flipping were used to expand the training samples due to the limited number of images in the dataset. To ensure that the experimental results only reflected the performance of the network structure and verify the effectiveness of the proposed method, the same data augmentation settings were used during the training process. The hyperparameter settings are shown in Table 2.