基于ResNet50与迁移学习的轮毂识别

针对人工进行轮毂分拣存在的误识别问题，采用一种基于ResNet50与迁移学习的神经网络模型来识别汽车轮毂。把预训练模型参数迁移到ResNet50卷积神经网络中，修改原网络的输出层，构建基于ResNet50的迁移学习模型，通过进一步训练轮毂数据集来微调模型参数，提取轮毂的细粒度特征。通过对比AlexNet、VGG11、VGG16与ResNet50模型在未使用微调、使用微调和冻结不同数量卷积层参数时的训练效率、准确率，证明ResNet50迁移学模型在冻结前7个Bottleneck残差块参数时不仅能缩短训练时间，并能在相同迭代周期下取得更高的准确率。在该冻结策略下训练生成TL-ResNet50迁移学习模型，分别对8种轮毂进行预测，得出每种轮毂的平均准确率达到99%以上。

Wheel Hub Recognition Based on ResNet50 and Transfer Learning

Aiming at the problem of false identification in manual wheel hub sorting, a neural network model based on ResNet50 and transfer learning is adopted to identify the model of automobile wheel hub.The parameters of the pretraining model are migrated to the ResNet50 convolution neural network, the output layer of original network is modified, and the transfer learning model based on ResNet50 is constructed.By comparing the training efficiency and accuracy of AlexNet, VGG11, VGG16 and ResNet50 when different volume convolution layer parameters are not fine-tuned, used fine-tuning and frozen, it is proved that the ResNet50 transfer model can not only shorten training time when the parameters of the seven bottleneck fragments are frozen but also achieve higher accuracy under the same iteration cycle.Under the freezing strategy, the TL-ResNet50 transfer learning model is trained to predict each of the eight hubs, and the average accuracy of each hub is over 99%.