基于深度学习的微观芯片字符识别系统

Design and research of chip micro character recognition system based on deep learning

针对传统的形态分割和模板匹配文字识别方法存在着识别精度低和不稳定的问题,为工业检测领域的芯片字符识别设计了一套基于深度学习的智能光学字符识别系统。该系统基于可微分二值化网络(detection with differentiable binarization network,DBNet)、方向分类器和卷积网络,3个阶段分别训练深度模型实现文本区域的检测、文本方向分类和字符识别,最后进行串联推理完成微观芯片字符的自动化识别。同时针对显微场景下芯片字符图像易受光照干扰,采用数据增强与扩充、更换网络骨架、更改网络卷积步长,解决了复杂背景下微观芯片字符识别易误检的问题。工业生产线上的实际测试结果表明,该系统的识别准确率达到99.9%,误检率3.4?,速度0.56 s/张,极大地提升了字符识别正确率和效率,降低了误检率。最终的识别结果可以直接在云端远程实时查看,简化了传统工业字符识别流程,有助于工业智能化检测进一步发展和提高。

From the perspective of industrial intelligent detection,,like morphological segmentation and template matching,this study provided a character recognition system based on deep learning for industrial chip character recognition.The system is based on detection with Differentiable Binarization Network(DBNet),direction classifier and convolution Network.The depth model is trained to detect text area,classify text direction and recognize characters in three stages.Serial reasoning was used to realize automatic recognition of microchip characters.At the same time,aiming at the fact that the chip character image is susceptible to illumination interference in the microscopic scene,data enhancement,data expansion and network structure optimization were carried out to solve the problem that the micro-chip character recognition is prone to false detection in the complex background.In practical tests on industrial production lines,the system achieved 99.9 percent accuracy,a rate of 3.4 errors per 10000 seconds,and a speed of 0.56 seconds per card,which significantly improved character recognition and efficiency and reduced false detection rates.The final recognition results can be directly viewed remotely and in real time in the cloud,which simplifies the traditional industrial character recognition process and contributes to the further development of industrial intelligent detection.