【摘要】 为提高芯片表面字符识别的实时性和准确率，提出一种基于EAST与SVTR的字符识别算法。针对EAST文本检测算法，将主干特征提取网络替换为轻量化的深度神经网络FasterNet-T0,减少网络的计算量；添加通道注意力机制自适应学习不同通道的权重分配，加强对重要特征的筛选。改进获得文本区域得分的损失函数，采用Dice损失缓解因图像背景面积过大导致误检的问题。文本方向校正算法对图像中任意方向的文本进行水平校正。由单一视觉模型的文本识别算法SVTR完成对字符的识别。实验结果表明，改进后文本检测算法的精确率、召回率较原算法分别提升了2.43%和4.66%,单帧图片的检测速度提升了0.005 s;添加文本方向校正算法后，识别准确率提升了1.92%。与现有方法对比，验证了该算法的有效性。更多还原

【Abstract】 To improve real-time performance and accuracy of chip surface character recognition, a character recognition algorithm based on EAST and SVTR was put forward. For the EAST text detection algorithm, the backbone was replaced by the lightweight deep neural network FasterNet-T0 to reduce the computational complexity of the network. A channel attention mechanism was added to adaptively learn the weight distribution of different channels to strengthen the recognition of important features. The loss function for obtaining text region scores was improved by adopting the Dice loss to alleviate the issue of false detection caused by overly large image backgrounds. The text direction correction algorithm was applied to enable the horizontal correction of text in images of any orientation, followed by character recognition, which was carried out by a single visual model for text recognition(SVTR). Experimental results indicate that, compared to the original algorithm, an increase of 2.43% in precision, 4.66% in recall, and 0.005 seconds in the detection speed of a single frame image for the improved text detection algorithm can be gained. An increase of 1.92% in recognition accuracy is brought about by the addition of the text direction correction algorithm. By comparing with existing method, the validity of the presented algorithm is demonstrated.更多还原