Abstract: Traditional convolutional neural networks use shallow-to-deep feature extraction to construct a classifier for image classification,which is thought to be easy to ignore shallow features.Based on the argon arc welding spot images collected in the industrial process,an en-AlexNet network was proposed by modifying Alexnet network.Batch normalization was used for data normalization,and meanwhile cross-connecting structure was introduced with its sensitivity analysis performed,by means of which the deep and shallow features were fused together.The Inception module was also embedded for multi-scale deep feature extraction.Softmax classifier was used to complete the classification.Results revealed that the average classification accuracy was 98.13%,which was better than both the traditional Support Vector Machine (SVM) method and typical convolutional neural networks.Moreover,the model was superior in aspects of size,convergence speed and average recall rate compared to typical convolutional neural networks.Good

Key words: improved AlexNet network, cross-connect structure, feature fusion, image classification, visual analysis, convolutional neural network, welding spot

摘要： 传统卷积神经网络是通过由浅入深的特征提取,构建分类器进行图像分类,该方式容易忽略浅层特征。对工业过程收集的氩弧焊点图像,基于AlexNet网络进行修改,提出一种增强的AlexNet(en-AlexNet)网络,该方法使用批归一化进行数据归一化,引入跨连结构同时进行灵敏性分析,并嵌入Inception模块进行多尺度深度特征提取,将深层特征与浅层特征进行融合,最后使用softmax分类器来判别焊点是否合格。最终平均分类准确度达到9813%,优于传统的支持向量机方法和其他的典型卷积神经网络模型。另外,模型的参数量、收敛速度和平均召回率均优于其他典型的卷积神经网络。结果验证了所提en-AlexNet网络在焊点的多类别分类上的有效性。

关键词: 改进AlexNet网络, 跨连结构, 特征融合图像分类, 可视化分析, 卷积神经网络, 焊接点