Yaguan Li

Rotation-invariant recognition of shapes is a common challenge in computer vision. Recent approaches have significantly improved the accuracy of rotation-invariant recognition by encoding the rotational invariance of shapes as hand-crafted image features and introducing deep neural networks. However, the methods based on pixels have too much redundant information, and the critical geometric information is prone to early leakage, resulting in weak rotation-invariant recognition of fine-grained shapes. In this paper, we reconsider the shape recognition problem from the perspective of contour points rather than pixels. We propose an anti-noise rotation-invariant convolution module based on contour geometric aware for fine-grained shape recognition. The module divides the shape contour into multiple local geometric regions(LGA), where we implement finer-grained rotation-invariant coding in terms of point topological relations. We provide a deep network composed of five such cascaded modules for classification and retrieval experiments. The results show that our method exhibits excellent performance in rotation-invariant recognition of fine-grained shapes. In addition, we demonstrate that our method is robust to contour noise and the rotation centers. The source code is available at https://github.com/zhenguonie/ANRICN_CGA.

Kinematics calibration for quadrupled robots is essential to ensuring motion accuracy and control stability. The angle of the leg joints of the quadruped robot is error-compensated to improve its position accuracy. This paper proposes a real-time high-precision kinematics calibration method for quadruped robots using machine vision and artificial neural networks to simplify the calibration process and improve calibration accuracy. The method includes two parts: identifying the markers fixed on the legs through target detection, calculating the center coordinates of the markers, and building an error model based on an artificial neural network to solve the angle error of each joint and compensate for it. A series of experiments have been carried out to verify the model’s accuracy. The experimental results show that, compared to traditional manual calibration, by adding an error correction model to the inverse kinematics neural network, the calibration efficiency can be significantly improved while the calibration accuracy is met.

Defect detection is a crucial direction of deep learning, which is suitable for industrial inspection of product quality in strip steel. As the strip steel production line continuously outputs products, it is necessary to take corresponding measures for the type of defect, once a subtle quality problem is found on steel strips. We propose a new defect area detection and classification method for automation strip steel defect detection. In order to eliminate the way of insufficient data in industrial production line scenarios, we design a transfer learning scheme to support the training of defect region detection. Subsequently, in order to achieve a more accurate classification of defect categories, we designed a deep learning model that integrated the detection results of defect regions and defects feature extraction. After applying our method to the test set and production line, we can achieve extremely high accuracy, reaching 87.11%, while meeting the production speed of the production line compared with other methods. The accuracy and speed of the model realize automatic quality monitoring in the manufacturing process of strip steel.

In the production of cold-rolled galvanized steel strips used for stamping car body parts, the in-situ and real-time defect detection is crucial for quality control, in which various types of defects inevitably occur. It is challenging to improve the accuracy of defect detection and classification by appropriate means to assist the manual screening process better. Defects under actual production conditions are often not prominent enough in defect characteristics, and there may be a significant similarity between different defect categories. To eliminate this weakness, we propose a data-driven deep learning approach named steel surface faulty detection attention net (SSFDANet) that uses images of the galvanized steel surfaces as input to identify whether the product is qualified and automatic classification of defect types instantaneously. This method can shorten product inspection time and improve the production line automation efficiency. In addition, the attention mechanism is utilized to enhance the performance of SSFDANet. Compared with the baseline ResNet, SSFDANet achieves a noticeable improvement in classification accuracy on test data. The well-trained model can successfully show an improved performance than the baseline models on the multiple types of faulty. Enhanced by SSFDANet with high classification accuracy, the defect rate of products is significantly reduced, and the production speed of the production line is significantly improved. Future prospective studies that are inspired by this article are also discussed.

Laser cladding is one of the critical technologies for additive manufacturing and rapid repair. Improving cladding performance by materials and process parameters is the leading research direction, but defects and instability of quality in the cladding process are inevitable. Therefore, it is necessary to study which factors are related to quality. In this paper, a new detection method is proposed to measure the radiation intensity of the reflected laser, laser scanning displacement, and temperature of the substrate while cladding. The characteristic values corresponding to the position of the cladding spots are extracted, the cladding quality is preliminarily evaluated and graded, and the correlation between them is verified with the method of machine learning nu-SVM. The results show that the accuracy of the model trained by 300 groups of data to predict the quality grades is 78.74%, which indicates that there is a strong correlation between these process variables and the cladding quality, and this method is feasible for the quality evaluation and control of the cladding process.

The increase of the spatial resolution in numerical computation always leads to a decrease in computing efficiency with respect to the constraint of mesh density. In response to this problem of the inability to perform numerical computation, we propose a novel method to boost the mesh-density in the finite element method (FEM) within 2D domains. Running on the von Mises stress fields of the 2D plane-strain problems computed by FEM, the proposed method utilizes a deep neural network named SMNet to learn a nonlinear mapping from low mesh-density to high mesh-density in stress fields and realizes the improvement of numerical computation accuracy and efficiency simultaneously. By introducing residual density blocks into SMNet, we can extract abundant local features and improve prediction capacity. The result indicates that SMNet can effectively increase the spatial resolution of stress fields under multiple scaling factors in mesh-density: 2 ×, 3 ×, and 4 ×. Compared with the targets, the relative error of SMNet is 1.67%, showing better performance than many other methods. SMNet can be generically used as an enhanced mesh-density boosting model of 2D physical fields for mesh-based numerical methods.