Target detection of electronic components on PCB (Printed circuit board) based vision is the core technology for 3C (Computer, Communication and Consumer Electronics) manufacturing companies to achieve quality control intelligent assembly robots. However, number large, shape different. At present, accuracy algorithm detecting all not high. This paper proposes an improved YOLO (you only look once) V3 (Version 3), which uses a real picture virtual with synthesized data as joint training...

ABSTRACT The detection of defects on industrial surfaces is essential for guaranteeing the quality and safety products. Deep learning‐based object methods have demonstrated impressive efficacy in applications recent years. However, due to complex variable shape defects, similarity between background, large intra‐class differences, small inter‐class differences lead low classification accuracy, it a great challenge achieve accurate defect detection. To overcome these challenges, this research...

Intelligent apple-picking robots can significantly improve the efficiency of apple picking, and realization fast accurate recognition localization apples is prerequisite foundation for operation picking robots. Existing methods primarily focus on object detection semantic segmentation techniques. However, these often suffer from errors when facing occlusion overlapping issues. Furthermore, few instance are also inefficient heavily dependent results. Therefore, this paper proposes an method...

Deep learning based on a convolutional neural network (CNN) has been successfully applied to stereo matching. Compared with the traditional method, speed and accuracy of this method have greatly improved. However, existing matching framework CNN often encounters two problems. First, many parameters, which leads running time being too long. Second, disparity estimation is inadequate in some regions where reflections, repeated textures, fine structures may lead ill-posed Through lightweight...

This paper addresses the robust enhancement problem in control of robot manipulators. A new hierarchical multiloop model predictive (MPC) scheme is proposed by combining an inverse dynamics-based feedback linearization and a nonlinear disturbance observer (NDO) based uncertainty compensation. By employing linearization, multi-link manipulator was decoupled to reduce computational burden compared with traditional MPC method. Moreover, NDO introduced into input torque signal compensate correct...

<abstract> <p>As an essential part of electronic component assembly, it is crucial to rapidly and accurately detect components. Therefore, a lightweight detection method based on knowledge distillation proposed in this study. First, student model was constructed. Then, we consider issues like the teacher student's differing expressions. A combination feature channel learn teacher's rich class-related inter-class difference features. Finally, comparative experiments were analyzed...

Due to nonlinearity and uncertainty of the robotic manipulator, design robot controller has a crucial impact on its performance motion trajectory tracking. In this paper, linear parameter varying (LPV) - model predictive (MPC) two-link manipulator is established then controller's optimal parameters are determined via newly developed meta-heuristic algorithm, transient search optimization (TSO). The proposed control method verified by set point nonlinear test set-point tracking, LPV-MPC...

Recent stereo matching methods, especially end-to-end deep networks, have achieved remarkable performance in the fields of autonomous driving and depth sensing. However, state-of-the-art algorithms, even with neural network framework, still difficulties at finding correct correspondences near-range regions object edge cues. To reinforce precision disparity prediction, present study, we propose a parallax attention algorithm based on improved group-wise correlation to learn content from...

2D raw image of Light Field (LF) camera needs to be decoded into 4D LF data for representation and processing. In decoding, the main lens is usually modeled as a thin while micro-lens array pinhole array. Although this model takes account distortion, it still difficult accurately characterize complex imaging relationship camera. order obtain more accurate data, paper proposes decoding method based on two-plane ray model. By calibrating corresponding pixel point camera, mapping between real...

To improve the accuracy of using deep neural networks to predict depth information a single image, we proposed an unsupervised convolutional network for single-image estimation. Firstly, is improved by introducing dense residual module into encoding and decoding structure. Secondly, optimized hybrid attention introduced network. Finally, stereo image used as training data realize end-to-end The experimental results on KITTI Cityscapes sets show that compared with some classical algorithms,...