Deep neural network (DNN) exhibits state-of-the-art performance in many fields including weld defect classification. However, there is still a large room for improving the classification performance over the generic DNN models. In this paper, a unified deep neural network with multi-level features is proposed for weld defect classification. Firstly, we define 11 weld defect features as inputs of our proposed classification model. Not limited to geometric and intensity features, 4 features based on the intensity contrast between weld defect and its background are proposed in this paper. Secondly, we construct a novel deep learning framework: a unified deep neural network, where multi-level features of each hidden layer are fused by the last hidden layer to predict the type of weld defect comprehensively. In addition, we investigate pre-training and fine-turning strategies to get better generalization performance with small dataset. Comparing with other classification methods like SVM and generic DNN model, our framework takes full advantage of multi-level features extracted from each hidden layer, an outstanding performance is shown where the classification accuracy is improved by 3.18% and 4.33% on the test dataset, to reach 91.36%.

深度神经网络（DNN）在包括焊缝缺陷分类在内的许多领域都表现出最先进的性能。但是，与通用DNN模型相比，仍有很大的改进空间。本文提出了一种具有多层次特征的统一的深度神经网络，用于焊接缺陷的分类。首先，我们定义11个焊接缺陷特征作为我们提出的分类模型的输入。不限于几何特征和强度特征，本文提出了基于焊接缺陷及其背景强度对比的四个特征。其次，我们构建了一个新颖的深度学习框架：统一的深度神经网络，其中每个隐藏层的多级特征与最后一个隐藏层融合在一起，以全面预测焊接缺陷的类型。此外，我们研究了预训练和微调策略，以利用较小的数据集获得更好的泛化性能。与其他分类方法（例如SVM和通用DNN模型）相比，我们的框架充分利用了从每个隐藏层提取的多级特征，显示了出色的性能，其中测试数据集的分类精度提高了3.18％和4.33％，达到91.36％。