FE-trans-net: Feature enhancement based single branch deep learning model for surface roughness detection

Abstract

Real-time surface roughness measurement is crucial for modern manufacturing. Machine vision-based roughness measurement algorithms have the advantages of non-contact, high efficiency and adaptability. Light source is important for online machine vision-based roughness measurement. And different materials may need different light sources to extract more features. From the previous study, dual light sources are better than single light source for surface roughness prediction. However, it is inconvenient to assign two light sources for online detection equipment, and the two-branch model structure for dual light sources doubles the size of the model. To fully utilize the advantages of dual light sources, and avoid its disadvantages, a feature enhancement based single branch deep learning model FE-Trans-Net is proposed in this paper for surface roughness detection. Firstly, a dataset of different materials (Cu, Zn, Al, and Fe) shined with different light sources (white, red, green, red laser and green laser) is collected. Two most suitable light sources for different materials are chosen for dual light sources model as offline pre-training. In this model, lightweight cross-attention fusion (CAF-Trans) and graph convolutional channel attention (GCC-Atten) modules are designed to fuse the feature information of different light sources from the spatial and channel perspectives respectively, enabling the single-branch model to acquire the feature learning capability of different light sources. Then a single-branch network layer is remained as the online prediction. With the feature enhancement, it preserves the advantages of dual light source images, and has only half size of the dual light source model. Extensive experiments validate the effectiveness and generalization performance of the CAF-Trans and GCC-Atten modules, and the proposed FE-Trans-Net network achieves the best accuracy-efficiency tradeoff compared to other popular deep neural networks.