基于改进MobileNetV2的钢板表面缺陷检测

Steel Plate Surface Defect Detection Based on Improved MobileNetV2

钢板表面缺陷伤痕类型多,行业数据公开率极低,训练样本不足使得深度学习难以应用于该领域。且MobileNetV2网络模型特征表示能力有限、鲁棒性较弱。针对上述问题,提出一种改进的MobileNetV2网络模型,可在小规模样本检测中拥有较高的准确率。重新设定网络模型中激活函数的上限,使模型更好地捕捉输入数据中的复杂模式和特征。提出一种新的瓶颈结构并减少网络层数,可以在通道维度上对特征图进行整合,提高模型的表示能力和特征提取能力。增强特征识别,提取更丰富和更具判别性的特征,提高模型的准确性和鲁棒性。实验结果表明,改进的MobileNetV2网络模型准确率高达98.7%,高于原网络和其他对比卷积神经网络,能有效检测小样本的钢板表面缺陷。

There are many types of defects on the surface of steel plates,the disclosure rate of industry data is very low,and the lack of training samples makes it difficult for deep learning to be applied in this field.The feature representation capability of MobileNetV2 network model is limited and its robustness is weak.To solve the above problems,an improved MobileNetV2 network model is proposed,which can have a high accuracy in small-scale sample detection.Resetting the upper limit of the activation function in the network model allows the model to better capture complex patterns and features in the input data.A new bottleneck structure is proposed and the number of network layers is reduced,which can integrate the feature graph in channel dimension and improve the representation and feature extraction ability of the model.Enhance feature recognition,extract more comprehensive and distinctive features,and enhance the accuracy and robustness of the model.Experimental results demonstrate that the improved MobileNetV2 network model achieves an accuracy of 98.7%,surpassing the original network as well as other convolutional neural networks.This improved model effectively detects surface defects in small steel plate samples.