基于改进CRNN的导弹编号识别算法研究

Research on Missile Number Recognition Based on Improved CRNN Algorithm

摘要:

海军某部队导弹出入库使用人工登计,导致了出入库过程浪费大量的人力时间。利用深度学习,可实现自动登记。在CRNN算法中加入非对称卷积,提升宽度感受区域,加入Attention机制对特征序列进行加权平均。通过在人工合成的数据集上进行实验对比分析,本文提出的改进CRNN算法识别目标导弹编号模型的准确率以及LOSS均达到较好的性能,较其他先进的文本识别算法,其字符准确率达到了98.9%,同时其平均编辑距离低至0.92,且经实际测试其均能准确识别出导弹编号。因此,利用改进的CRNN算法识别导弹编号,辅助工作人员进行导弹出入库自动登记方案是可行的。

Abstract:

A certain naval force used manual check-in for missiles in and out of the warehouse, which caused a lot of manpower and time to be wasted during the warehouse. Using deep learning, automatic registration can be achieved. By adding asymmetric convolution to the CRNN algorithm, the width of the perception area is increased, and the Attention mechanism is added to perform a weighted average of the feature sequence. Through experimental comparative analysis on artificially synthesized data sets, the accuracy and the loss of the target missile number model based on the improved CRNN algorithm have achieved better performance. Compared with other advanced text recognition algorithms, its character accuracy has reached 98.9%, and its average editing distance is as low as 0.92. And it can accurately identify the missile number after actual testing. Therefore, it is feasible to use the improved CRNN algorithm to identify the missile number and assist the staff in the automatic registration of missiles in and out of the warehouse.