基于改进 SqueezeNet 的棒状物表面缺陷识别

高速流水线生产的棒状物极易产生各种表面缺陷，但基于传统图像处理的缺陷识别方法易受环境影响、可靠性低，而基于深度学习的缺陷识别方法存在模型过大、识别准确率受制于样本数量等问题。因此，本文提出了一种基于改进SqueezeNet的棒状物表面缺陷识别系统。设计了可获取圆周对称小体积棒状物全表面图像的采集装置，并在轻量级卷积神经网络SqueezeNet中引入注意力模块以改善模型的特征提取效果，利用数据平衡方法提升数据集内少数类样本的识别准确率，利用迁移学习的方法进行深度学习训练，减轻数据集样本不足对训练效果的影响。以生产线上的卷烟烟支为研究对象，采集其圆周表面图像进行实验，结果表明，改进方法在少样本条件下的分类准确率达到了94.49%,其中对于少数类样本的F1分数提高了31.19%,单张图像检测时间约1.66 ms,模型轻量化，可满足工业生产线中棒状物实时缺陷识别的需求。

Rods surface defect identification based on improved SqueezeNet

The rods produced by the high-speed assembly line are highly susceptible to various surface defect, but the defect identification method based on conventional image processing is unreliable and susceptible to environmental factors, while the defect identification method based on deep learning suffers from oversized models and recognition accuracy that is constrained by the quantity of samples. Therefore, this paper suggests an identification system of rods surface defect based on improved SqueezeNet. An acquisition device was designed to obtain the full surface image of the circumferential rods, and the attention module is introduced into the lightweight convolutional neural network SqueezeNet to improve the feature extraction effect of the model, data balancing is used to improve the recognition accuracy of minority samples, transfer learning is employed for deep learning training to minimize the impact of insufficient samples on the training effect. Taking the cigarette on the production line as the research object, the circumferential surface image of the cigarette is collected for experiment, the results show that the classification accuracy of the improved method under the condition of few samples reaches 94. 49%, with the F1 score for minority samples being improved by 31. 19%, and the detection time of a single image being approximately 1. 66 ms. Additionally, the model is lightweight, meeting the need for rods in industrial production lines to have real-time defects recognized.