Abstract: Based on an improved  spatial pyramid attention module, we  enhanced the performance of lightweight networks in fine-grained image classification tasks. By combining global and local features, the  improved model enhanced the classification performance of lightweight networks without significantly increasing the number of parameters. The experimental results  on the Stanford Dogs dataset show  that the lightweight network equipped with this module significantly improves accuracy, even surpassing some classical models. This method expands the application scope of lightweight networks on resource-constrained devices and provides an efficient and low-computational-cost solution for fine-grained image classification problems.

Key words:  , image classification, fine-grained classification, attention mechanism, lightweight network