AMS-Net: Attention mechanism based multi-size dual light source network for surface roughness prediction

Real-time and efficient surface roughness measurement is very essential for machining. Previously proposed non-contact roughness prediction methods usually use surface roughness images taken by a single light source. However, different light sources have a different reflection feature on the same roughness surface. A novel surface roughness prediction method of deep learning network model is proposed in this paper which is based on channel and spatial attention mechanisms, and embedded into the multi-size parallel framework (AMS-Net) for dual light sources. The network uses two branches which extract the features of surface roughness implied in images taken under different light sources. In each branch, the multi-size parallel convolution module (MSP) is constructed to extract parallel imagery pertaining to multi-size feature information of images with different roughness. In order to better fuse the surface roughness images taken by the dual light sources, the dual light source spatial attention module (DSA) and dual light source channel attention module (DCA) are proposed to project the space and channel respectively, while interacting with the roughness feature map from the MSP module. The dual light source roughness data set is built by white light and red laser for microimaging in this paper. Comparison experiments are implemented with other popular classification deep learning models. The results of experiment show the proposed novel AMS-Net network that achieves the best classification accuracy.