The surface defects of the workpiece affect the workpiece quality. In order to detect workpiece surface defects more accurately, an automatic detection convolutional neural networks-based method is proposed in this paper. Firstly, a convolution network classification model (SCN) with symmetric modules is proposed, which is used as backbone of our method to extract features. And then, three convolution branches with FPN structure are used to identify the features. Finally, an optimized IOU (XIoU) is designed to define the loss function, which is used for detection model training. In addition to the public datasets NEU-CLS and NEU-DET, a classification dataset and a detection dataset of surface defects on hearth of raw aluminum casting are established to train and evaluate our model. On the basis above, the proposed backbone SCN was compared with Darknet-53 and ResNet-101 to present its superiority in classification performance. The average accuracy of SCN on NEU-CLS and self-made data sets are 99.61% and 95.84% respectively, which is significantly higher than the other two classification models. Then, in order to show the effectiveness and superiority of the proposed automatic detection method, the detection performance of the method is compared with the Faster-RCNN series and the YOLOv3 series. The result shows that our model achieves 79.89% mAP on NEU-DET and 78.44% mAP on self-made detection dataset. Our model can detect at 23f/s when the input image size is 416 × 416 × 3. The detection performance of our model is significantly better than other models. The results show that the proposed method has better performance and can be used for real-time automatic detection of workpiece surface defects.

工件的表面缺陷会影响工件质量。为了更准确地检测工件表面缺陷，提出了一种基于卷积神经网络的自动检测方法。首先，提出了一种具有对称模块的卷积网络分类模型（SCN），该模型被用作我们的特征提取方法的骨干。然后，使用三个具有FPN结构的卷积分支来识别特征。最后，设计了优化的IOU（XIoU）来定义损失函数，该函数用于检测模型训练。除了公共数据集NEU-CLS和NEU-DET，还建立了分类数据集和生铝铸件炉膛表面缺陷检测数据集来训练和评估我们的模型。在上述基础上，将该提议的主干SCN与Darknet-53和ResNet-101进行比较，以显示其在分类性能方面的优越性。NEU-CLS和自制数据集上SCN的平均准确度分别为99.61％和95.84％，显着高于其他两个分类模型。然后，为了展示所提出的自动检测方法的有效性和优越性，将该方法的检测性能与Faster-RCNN系列和YOLOv3系列进行了比较。结果表明，该模型在NEU-DET上的mAP达到79.89％，在自制检测数据集上的mAP达到78.44％。当输入图像尺寸为416×416×3时，我们的模型可以以23f / s的速度进行检测。我们模型的检测性能明显优于其他模型。