ISD-YOLO: A lightweight and efficient feature extraction network for invoice seal detection

Invoice seal detection plays an important role in intelligent invoice processing by improving review efficiency and reducing potential financial risks. However, existing methods still face challenges in multi-scale feature representation, complex background interference, and balancing detection accuracy with computational efficiency. In this study, a lightweight detection model, ISD-YOLO, is proposed for invoice seal detection. Specifically, the C3k2-ConvFormer-CGLU module (C3k2_CFC) is designed to enhance multi-scale feature representation by combining local feature enhancement and gated feature selection. A C2BRA module based on bi-level routing attention is introduced to improve the extraction of discriminative seal features by dynamically selecting important regions. In addition, a deformable attention module (DAttention) is incorporated to enhance the focus on key regions and reduce background interference. Furthermore, a lightweight detection head (EfficientHead) is developed to reduce computational complexity while maintaining detection performance, providing potential for resource-constrained applications. Experiments were conducted on a dataset consisting of public seal images and real invoice images. Compared with the baseline model, ISD-YOLO improves mAP50 and mAP50:95 by 2.1% and 2.3%, respectively. Compared with existing detection methods, ISD-YOLO achieves an F1-score of 95.3% and an mAP50 of 97.2%, while requiring only 2.3 M parameters and 5.0 GFLOPs. The results demonstrate that ISD-YOLO achieves a favorable balance between detection accuracy and computational efficiency under the evaluated invoice seal detection conditions. Its lightweight design indicates potential for resource-constrained applications, while further validation under broader scenarios and on actual edge devices is still required.