基于深度学习的铸件字符识别

Casting Character Recognition Based on Deep Learning

针对铸件字符人工识别效率低、人工记录易出错,现有的字符识别方法无法应对工业场景下复杂的铸件字符,且场景本身存在极端光照、遮挡、模糊的问题,提出了改进的PGNet网络。该网络在识别水平文本的同时,也能很好地识别弯曲文本和不规则文本。针对铸件字符数据量不足的情况,加入STN矫正模块进行数据增强,不同的实验结果表明,对准召的提升大于1%。此外,通过优化PGNet网络的损失函数,降低了误识别率。通过对PGNet网络的改进,在一定程度上解决了上述问题,使得铸件字符的溯源与管控过程更加准确和高效。

In view of the low efficiency of casting character recognition, easy error in manual recording, the existing character recognition methods can not deal with the complex casting characters in industrial scenes, and the scene itself has extreme lighting, blocking, fuzzy problems, an im-proved network has been proposed. This network can recognize not only horizontal text but also curved text and irregular text. In view of the insufficient amount of casting character data, the correction module was added for data enhancement. Different experimental results show that the increase of alignment is greater than 1%. In addition, by optimizing the loss function of PGNet network, the misidentification rate is reduced. Through the improvement of PGNet network, the above prob-lems are solved to a certain extent, making the casting character tracing and control process more accurate and efficient.