Due to the random surface texture and weak features of the grinding processing, most current visual measurement of grinding surface roughness is predicted by designing feature indicators, but its imaging environment is more demanding and the indicator design is more artificial. While deep learning can achieve feature self-extraction, the existing deep learning-based feature extraction uses a single convolutional model to extract features, which tends to make the extracted features noise, with low resolution and poor perception of details. To address the above problems, this paper proposes a feature fusion-based method for measuring grinding surface roughness. The method adopts different feature extraction ways and fuses the extracted features to obtain more features while also improving the generalization ability of the model, and verifies the adaptability of the model in different lighting environments. The experimental results show that deep learning self-extracted features based on feature fusion can effectively solve the problem of weak feature information on grinding surface roughness that is difficult to identify, and the model has high detection accuracy across different lighting environments, thus laying the foundation for the automated visual online measurement of grinding surface roughness.

Color image indices used to characterize surface roughness are more sensitive than grayscale images or spectral indices owing to the rich information contained in color images. In this paper, a method for measuring grinding surface roughness grinding based on the singular value entropy of the color image quaternion matrix is proposed. Color images captured at different levels of surface roughness are analyzed using a quaternion matrix, which is subjected to singular value decomposition from which a quaternion singular value entropy is derived as an index for evaluating roughness. Our experimental results show that this method for measuring grinding surface roughness based on quaternion singular value entropy is a more feasible roughness detection method than color difference indices because the singular value entropy is more strongly correlated with the actual roughness, with the monotonicity of the entropy decreasing more significantly as the roughness increases. Roughness prediction results obtained using a support vector machine also support the feasibility of measuring surface roughness based on the singular value entropy of the color image quaternion matrix, which can provide a reliable engineering application for the automatic measurement of surface roughness. Finally, the high degree of correspondence between the pure quaternion matrix and the image color matrix in the mathematical structure provides a broad mathematical space for the design and optimization of the color index.

In a realistic scenario where a large number of workpieces need to be measured, any measurement method that can detect roughness only for a single workpiece is very limited in terms of measurement efficiency. To address this problem, a multi-object surface roughness detection model based on Faster R-CNN is proposed in this paper. The model features milled workpiece images with a convolutional neural network. And the obtained features will feed into a Region Proposal Network for inferring those regions where workpieces may be present. The regions and features go through a ROI pooling layer and a predictor to get more accurate target regions and measure the roughness of the workpieces in the regions. The experimental results show that the model proposed in this paper can accurately detect those regions where workpieces are present in the image and detect the corresponding roughness grade of the workpieces. A mean average precision of 97.80% and a detection speed of 5.82 fps for the test set of milled workpieces were achieved by the model under different placement angles and variable light conditions.

Current machine vision methods for surface measurement rely excessively on feature design to quantify surface morphology and build predictive models, but metric design suffers from human intervention, and data acquisition is heavily dependent on light source environment and shooting angle. This paper uses the Swin-Transformer model to evaluate and classify roughness directly on colour images of milled samples acquired in a convertible image environment. The images used in the experiment are taken in a completely dark environment and in an environment disturbed by light sources, such as LED energy-saving lamps during the day. By using two kinds of lenses and combining custom light sources and ordinary light sources, respectively, and for different angles, it deeply simulates the environment of online inspection of industrial production. The roughness classification results prove that the method has very good robustness to light source environment and shooting angle, avoiding the artificial design and extraction of image features. The accuracy of the validation set of samples can reach 98.94%, while the accuracy of the test set can reach 97.54%. To wrap it up, the method provides an optimised strategy for visual roughness measurement in industrial production.

Existing roughness machine vision inspection methods based on designed indices require the artificial construction of features associated with roughness based on image information. Such methods are less universal but highly demanding for the imaging environment and cannot be directly applied to general industrial environments. The roughness machine vision inspection methods based on traditional deep learning, on the other hand, rely heavily on a large number of training samples, and the model needs to be retrained when the recognition of new categories is required by the inspection task. Therefore, a machine vision inspection method for milling surface roughness grade based on the improved model-agnostic meta-learning (MAML++) network was proposed to address the above problems. This method could not only automatically extract the image features associated with roughness but also recognize the surface roughness grade of milling workpieces by a small number of labeled samples. It reduced the need for sample size and can recognize new categories without retraining the model. The experimental results showed that the method exhibited high accuracy in roughness inspection tasks with different sample distributions and had promising robustness to complex light environments. In addition, in terms of the requirement for recognizing new roughness categories, a low-cost and fast roughness inspection task could be accomplished by acquiring fewer labeled samples. In summary, the proposed model has the potential to be a fast inspection system that can be applied to surface roughness inspection for most industrial applications. It provides the technology for inline inspection of machine vision roughness in general industrial lighting environments.