基于机器视觉的多种类标签识别方法研究

Research on Multi-type Label Recognition Method Based on Machine Vision

针对目前印刷标签复杂多样难以识别和分类,以及各种缺陷造成识别准确率低的问题,提出了一种基于机器视觉的多种类标签检测方法。首先搭建由相机、镜头和光源组成的采样平台,将各种类标签采集后用作模型训练的数据集,然后利用最小外接矩形并稳健回归的方式对图像进行畸变矫正,通过Laplacian算子、高斯滤波算法、Otsu算法消除噪声产生的影响,最后建立了一个改进的CRNN+CTC网络结构模型,其中加入BN算法和Adam算法提高模型的泛化能力和收敛速度,使用双向BLSTM网络减小梯度消失或爆炸,再加入CTC损失函数实现输入数据与给定标签的对齐问题。实验结果表明,改进后的方法相较于传统分割字符算法,识别准确率提升至98.2%;相较于原CRNN+CTC算法,识别速度提升至37 ms/张,达到了工业使用需求。

Aiming at the problems of complex and diverse printed labels that are difficult to identify and classify,as well as the low recognition accuracy caused by various defects,a multi-type label detection method based on machine vision was proposed.Firstly,a sampling platform composed of camera,lens and light source was built,and various labels were collected and used as data sets for model training.Then,the image distortion was corrected by using the minimum external rectangle and robust regression,and the influence of noise was eliminated by Laplacian operator,Gaussian filtering algorithm and Otsu algorithm.Finally,an improved CRNN+CTC network structure model was established,in which BN algorithm and Adam algorithm were added to improve the generalization ability and convergence speed of the model,bidirectional BLSTM network was used to reduce the gradient disappearance or explosion,and CTC loss function was added to realize the alignment problem between input data and given labels.Experimental results show that the recognition accuracy of the improved method is 98.2%compared with the traditional character segmentation algorithm.Compared with the original CRNN+CTC algorithm,the recognition speed is increased to37 ms/piece,which meets the demand of industrial use.