基于注意力机制的接线端子文本检测与识别

Text Detection and Recognition of Terminal Blocks Based on Attention Mechanism

针对当下变电站二次保护屏柜电缆接线仍采用传统人工验收方式,存在效率低、容易漏检、错检等问题,提出一种基于注意力机制的接线端子文本检测与识别方法。在文本检测阶段,针对接线端子弯曲倾斜、排列密集等问题,提出一种改进DBNet方法。使用SwinTransformer提取图像基础特征,搭建特征金字塔网络,提取并融合多尺度的图像特征,输出连接SEblock,增强重要特征权重,使检测框定位更加精准。在文本识别阶段,提出一种改进CRNN方法,使用ResNet提取特征,并在残差模块中加入SEblock,强化重要通道特征,进一步提升识别准确率。在检测和识别数据集上分别进行验证,结果表明:在检测数据集中,改进DBNet的精准率为95.6%,召回率为82.9%,调和平均数达到88.8%;在识别数据集中,改进CRNN方法的字符识别准确率达到87.2%。

At current,the cable wiring of secondary protection screen cabinet in substation still adopts traditional manual acceptance method,and there are problems such as low efficiency,easy to miss and error check.a text detection and identification method of wiring terminal based on attention mechanism is proposed.To solve this problem,a text detection and recognition method based on the attention mechanism is proposed.In the text detection stage,aiming at the problems of bending,inclination and dense arrangement of terminals,an improved DBNet method is proposed.Swin Transformer is used to extract basic image features,a feature pyramid network is built to extract and fuse multi-scale image features,and the output is connected to SE block to enhance Important feature weights make the detection frame positioning more accurate.In the text recognition stage,an improved CRNN method is proposed,using ResNet to extract features,and SE block is added into the residual module to strengthen the important channel features and further improve the recognition accuracy.Validation is performed separately on the detection and recognition data sets,the experimental results show that the precision of the improved DBNet in the detection data set is 95.6%,the recall is 82.9%,and the Hmean is 88.8%;the character recognition accuracy of the improved CRNN method in the recognition data set reaches 87.2%.