Multilevel Deep Feature Fusion for Surface Roughness Detection of Milled Workpieces

• In order to solve the problem of the need for rapid roughness measurement of workpieces and the difficulty of existing methods in achieving rapid online detection. • This paper built a dual branch multilevel feature information fusion network (including a CNN branch and a ViT branch) for the task of measuring the surface roughness of milled workpieces. • Under the same experimental parameter configuration, ablation experiments results showed that the model proposed in this paper has faster convergence speed and higher accuracy. • Comparative experiments were conducted with mainstream models. The results show that the accuracy of the model proposed in this paper was the highest, reaching 99.92%. Existing machine vision-based methods for measuring workpiece surface roughness require manual design of roughness-related feature indicators and have stringent imaging environment requirements, making them less suitable for direct application in manufacturing processes. On the other hand, deep learning-based methods rely on large training datasets and often require retraining when the measurement target changes. To address these issues, a dual-branch multilevel deep feature fusion model based on a convolutional neural network (CNN) and a Vision Transformer (DeiT-S) is proposed for workpiece surface roughness measurement. Feature fusion is performed at both intermediate and final stages of the two branches, enabling the model to simultaneously capture fine-grained local texture patterns related to tool marks and global contextual information related to illumination variations and long-range surface patterns, both of which are critical for accurate surface roughness classification. Experimental results show that with a limited dataset of 7,571 samples, the proposed model achieves a classification accuracy of 99.92% (±0.05%), outperforming single-branch ResNet18 and DeiT-S by 1.24% and 2.05%, respectively. The key innovations include (1) a dual-branch feature fusion mechanism that jointly models local textures and global illumination dependencies and (2) enhanced fine-grained feature extraction using Convolutional Block Attention Modules (CBAM), which effectively reduces the dependence on large-scale training data. Furthermore, computational efficiency analysis demonstrates that the proposed model achieves real-time inference performance, making it suitable for fast online roughness inspection systems.