基于轻量神经网络的钢材表面缺陷快速识别

为满足生产环境中钢材表面缺陷实时分类识别的需求,提出一种可部署于移动端的轻量卷积神经网络模型用于高效识别钢材表面缺陷。通过知识蒸馏将ResNet50和MobileNetV3分别作为教师模型和学生模型,生成改进的轻量神经网络模型MobileNetV3_small_tp。利用MobileNetV3_small_tp作为预训练模型,选用NEU带钢表面缺陷数据集进行微调,同时使用数据增强以模拟实际工况,得到模型在测试集中精度达到100%。最后将模型部署于手机上进行实际钢材表面缺陷识别,移动端模型测试、验证以及实际检测结果均显示MobileNetV3_small_tp模型具有流畅的运行速度和较高的识别率,能够实现现场钢材表面缺陷的准确高效识别。

Fast Identification of Steel Surface Defects based on Lightweight Neural Network

In order to meet the needs of real-time classification and recognition of steel surface defects in the production environment, this paper proposes a lightweight convolutional neural network model that can be deployed on mobile devices to efficiently identify steel surface defects. Through knowledge distillation, ResNet50 and MobileNetV3 are used as teacher model and student model respectively to generate an improved lightweight neural network model MobileNetV3_small_tp. The MobileNetV3_small_tp model is then pre-trained on the NEU strip steel surface defect dataset, which is fine-tuned using data enhancement to simulate actual working conditions. As a result, the accuracy of the obtained model in the test set reaches 100%. Finally, the model was deployed on a mobile phone to identify actual steel surface defects. The mobile device model test, verification and actual inspection results all show that the MobileNetV3_small_tp model has a smooth running speed and a high recognition rate, which can achieve accurate and efficient identification of on-site steel surface defects.