Efficient banknote denomination recognition using a lightweight CNN with enhanced channel attention

Automatic banknote recognition systems play a significant role in financial automation processing and assistive applications for visually impaired individuals. To address the challenges in fine-grained Renminbi (RMB) image recognition, including subtle inter-class differences, high local texture similarity, and the limited representation capability of lightweight models, this study proposes a lightweight convolutional neural network based on the Max-Avg Fusion Channel Attention (MAFCA) mechanism. By integrating global average pooling and global max pooling information, MAFCA enhances channel feature modeling and improves the network’s capability to capture discriminative fine-grained features. Based on this design, the proposed module is embedded into a lightweight convolutional architecture and evaluated on a dataset containing ten categories of RMB images. Experimental results demonstrate that the proposed model achieves stable performance improvements under five-fold cross-validation and outperforms both the baseline model and the Squeeze-and-Excitation (SE) attention-based model. Furthermore, the proposed approach effectively enhances classification performance while maintaining low computational overhead, thereby validating the effectiveness and robustness of MAFCA in fine-grained image recognition tasks.