基于Mask R-CNN的试管-支架系统Data Matrix码识别方法-《福建工程学院学报》

文章信息/Info

Title:: A Mask RCNN based method for Data Matrix code recognition from tube-rack system

Author(s):: LIU Shijian1; LIN Jinjia1; CHEN Zican1; ZOU Zheng; Fujian Provincial Key Laboratory of Big Data Mining and Applications

Keywords:: tuberack system; Mask R-CNN; Data Matrix code; data synthesis; laboratory automation

摘要:: 在试管-支架自动化系统的输入图像中,Data Matrix(DM)码呈现为多个小目标,图像存在成像模糊、边缘干扰严重等问题,使得传统方法难以达到良好的识别效果。为此,提出一种基于深度学习的Data Matrix码识别方法DeepDMCode,以Mask R?CNN模型为基础,通过内容差异化数据合成和同步自动化标注,实现训练数据的增强,提升模型的学习能力。在模型分割结果的基础上,提出一种旋转校正方法,确保可用标准解码库实现DM 码的解码。以分辨率为1600×1200、支架容量为96的数据实验表明,由于该方法在前期码定位阶段最大程度地还原码边界信息,准确度可达0.92(mIoU),完成单张图像中所有DM识别的平均速度为5.2s,优于YOLO、SegNet、CenterNet等主流工业基准算法。

Abstract:: The Data Matrix (DM) codes from the tuberack automatic system can be seen as multiple small targets in the input image, which has problems such as blurred imaging and serious edge interference and makes it difficult for traditional methods to achieve good recognition results. Therefore, a Data Matrix code recognition method based on deep learning named DeepDMCode was proposed. Based on the Mask RCNN model, the training data was enhanced and the learning ability of the model was improved through contentdifferentiated data synthesis and synchronous automatic annotation. Based on the segmentation results of the model, a rotation correction method was proposed to ensure that the DM code can be decoded using a standard decoding library. Experiments were carried out with images of resolution 1 600×1 200, which were captured from racks of 96 capacity. Results show that the method can restore the code boundary information to the greatest extent in the early code positioning stage, and the accuracy can reach 0.92 ( mIoU ). The average speed of completing all DM recognition in a single image is 5.2 s, which is better than mainstream industrial benchmark algorithms such as YOLO, SegNet, and CenterNet.