Roughness detection based on multiple lighting conditions and deep learning

In surface roughness measurement, traditional machine vision detection methods are based on human-designed image feature indicators, which requires to manual intervention and have low efficiency. Therefore, this paper presents a method for measuring surface roughness based on a convolutional neural network. This method utilizes ResNet50 to automatically extract multi-scale image features. Through the bidirectional feature fusion module, different scale feature information is fused. Finally, processing the rich feature information after fusion through fully connected layers regresses accurate roughness values. In addition, the influence of lighting conditions on roughness measurement is analyzed, such as the wavelength and incidence angle of the laser. In our self-generated dataset, the best lighting condition is achieved using the red laser with a wavelength of 658 nm and an angle of incidence of 60°. The corresponding neural network model achieved an average accuracy of $95.9 \%$ on the test set.