Model performance comparison.

Recognition of the key text of the Chinese seal can speed up the approval of documents, and improve the office efficiency of enterprises or government administrative departments. Due to image blurring and occlusion, the accuracy of Chinese seal recognition is low. In addition, the real dataset is very limited. In order to solve these problems, we improve the differentiable binarization detection algorithm (DBnet) to construct a model DB-ECA for text region detection, and propose a model named LSTR (Lightweight Seal Text Recognition) for text recognition. The efficient channel attention module is added to the differentiable binarization network to solve the feature pyramid conflict, and the convolutional layer network structure is improved to delay downsampling for reducing semantic feature loss. LSTR uses a lightweight CNN more suitable for small-sample generalization, and dynamically fuses positional and visual information through a self-attention-based inference layer to predict the label distribution of feature sequences in parallel. The inference layer not only solves the weak discriminative power of CNN in the shallow layer, but also facilitates CTC (Connectionist Temporal Classification) to accurately align the feature region with the target character. Experiments on the homemade dataset in this paper, DB-ECA compared with the other five commonly used detection models, the precision, recall, F-measure are the best effect of 90.29, 85.17, 87.65, respectively. LSTR compared with the other five kinds of recognition models in the last three years, to achieve the highest effect of accuracy 91.29%, and has the advantages of a small number of parameters and fast inference. The experimental results fully prove the innovation and effectiveness of our model.中文印章关键文本识别可加速公文审批进程，提升企业或政府行政部门的办公效率。受图像模糊与遮挡因素影响，当前中文印章识别的准确率偏低；此外，真实可用的印章数据集极为有限。为解决上述问题，我们对可微分二值化检测算法（Differentiable Binarization, DBnet）进行改进，构建了用于文本区域检测的DB-ECA模型，并提出一款名为LSTR（Lightweight Seal Text Recognition，轻量级印章文本识别）的文本识别模型。我们在可微分二值化网络中加入高效通道注意力模块（Efficient Channel Attention, ECA）以解决特征金字塔冲突问题，并改进卷积层网络结构，延迟下采样操作以减少语义特征损失。LSTR采用更适配小样本泛化的轻量级卷积神经网络（Convolutional Neural Network, CNN），并通过基于自注意力的推理层动态融合位置信息与视觉信息，并行预测特征序列的标签分布。该推理层不仅解决了卷积神经网络在浅层特征上判别力较弱的问题，还便于CTC（Connectionist Temporal Classification，连接主义时序分类）将特征区域与目标字符精准对齐。本文在自制数据集上开展对比实验：相较于其他五种常用检测模型，DB-ECA的精确率、召回率与F-measure（F度量）分别达到90.29、85.17、87.65，各项指标均为最优。相较于近三年提出的其他五种识别模型，LSTR实现了最高91.29%的识别准确率，同时具备参数量少、推理速度快的优势。实验结果充分验证了所提模型的创新性与有效性。