Changeable environment visual detection of grinding surface roughness based on lightweight network

The texture features of the grinding surface are not obvious, when the light source environment and shooting angle change, the prediction model of grinding surface roughness based on the deep learning method is greatly affected. Moreover, when applied to the mobile end or the embedded end, the model requires smaller size and faster inference time. This paper proposes a surface roughness detection model for grinding based on lightweight network MoCoViT, which can be applied to changeable imaging environments. This model uses the advantages of Transformer and mobile convolutional networks to improve performance and efficiency, so as to complete the recognition of grinding surface roughness in a transformable image environment. For different angles and different light sources, three experiments are designed to test the recognition effect of the model under unknown angles and unknown lighting conditions. The experimental results show that the grinding surface roughness detection model based on MoCoViT network has the highest accuracy in the test set. It proves that MoCoViT model has good robustness to the grinding surface roughness detection in the image changeable environment, which lays a foundation for the industrial application of online visual detection of grinding surface roughness.