A Lightweight Network for the Hot-Rolled Steel Strip Surface Defect Detection

With the amazing development of deep learning, more and more research apply it to the field of computer vision in recent years. However, most object detection networks are difficult to apply in our real life because of the complex structures and numerous parameters. So, it is meaningful to focus our attention on designing some lightweight network architecture in the future. In this paper, we will propose a new convolutional neural network that named M2I3-SSD (MobileNetV2-InceptionV3 Single Shot MultiBox Detector) and its variant named M2I3R-SSD (MobileNetV2-InceptionV3-NoRelu6 Single Shot MultiBox Detector) to detect the surface defection of the hot-rolled steel strip. Then fine-tune as a transfer learning method was used in our paper to reduce the overfitting problem caused by small database. And we do a lot of experiments to verify the performance about the proposed model in the end.