Method for Qin Bamboo Slip Text Detection Based on an Enhanced DBNet Model

An improved DBNet-based detection method is proposed to address the detection challenges posed by background interference, texture breakage, and dense arrangements in the text of unearthed Qin bamboo slips. This approach incorporates a text collaborative learning module and an adaptive weighted feature pyramid into the DBNet framework. ResNet18, augmented with deformable convolution and the text collaborative learning module, is utilized during the feature extraction stage. Features of characters and blank areas are collaboratively sampled through a convolutional cascade with varying receptive fields, effectively mitigating the issue of texture breakage. The enhanced adaptive weighted feature pyramid optimizes the feature map’s receptive field and attention fusion, improving the model’s ability to learn complex backgrounds and densely arranged text. Furthermore, the modified loss function applies bilateral upsampling to improve the adaptability of the binarization process. Experimental results on the synthetic Qin bamboo slips text detection dataset demonstrate that the proposed method achieves an accuracy of 99.82%, with F -score and recall rates of 97.02% and 94.38%, respectively-marking improvements of 1.35% and 2.05% over the original DBNet model. These results highlight the method’s superior performance and balanced effectiveness.