MGCNet: A Multi-scale Grouped Convolution-Based Seal Detection Method for Painting and Calligraphy Works

With the advancement of digital culture, there is a continuous integration of deep learning and artwork. This paper proposes a multi-scale grouped convolutional model, abbreviated as MGCNet, to address the time-consuming and laborious issue of traditional seal detection. Firstly, the multi-scale grouping convolution is constructed based on the principles of grouping and multi-scale ideas, allowing for the extraction of features with varying semantic information. Secondly, the convolution module in Bottleneck is replaced within the YOLOv8 feature extraction layer, resulting in a reduction of parameters in the model without compromising detection accuracy. Additionally, Mixed Local Channel Attention (MLCA) is introduced to further optimize the number of parameters in the MGCNet network model. Finally, an analysis is conducted on both the Self-constructed Painting and Calligraphy seal dataset and the PASCAL VOC2012 dataset. The method is compared with existing convolutional modules such as PConv, DWConv, and SCConv focusing on seal region detection accuracy, computational complexity, number of parameters, and model weight file size. The experimental results reveal that compared to YOLOv8, the proposed method achieves an 11.3% reduction in the number of parameters on the PASCAL VOC2012 dataset while enhancing accuracy by 0.7%. This not only reduces storage costs but also facilitates deployment in resource-constrained environments such as mobile devices.