To create an intelligent surface region of interests (ROI) 3D quantitative inspection strategy a reality in the continuous casting (CC) production line, an improved 3D laser image scanning system (3D-LDS) was established based on binocular imaging and deep-learning techniques. In 3D-LDS, firstly, to meet the requirements of the industrial application, the CCD laser image scanning method was optimized in high-temperature experiments and secondly, we proposed a novel region proposal method based on 3D ROI initial depth location for effectively suppressing redundant candidate bounding boxes generated by pseudo-defects in a real-time inspection process. Thirdly, a novel two-step defects inspection strategy was presented by devising a fusion deep CNN model which combined fully connected networks (for defects classification/recognition) and fully convolutional networks (for defects delineation). The 3D-LDS’ dichotomous inspection method of defects classification and delineation processes are helpful in understanding and addressing challenges for defects inspection in CC product surfaces. The applicability of the presented methods is mainly tied to the surface quality inspection for slab, strip and billet products.

As an important mechanical transmission device, the camshafts with highly reflective surface (HRS) are widely used in various automobiles. The automobile manufacturers have increasingly strict requirements for the processing quality of camshafts. However, affected by multiple processing steps, it is impossible to achieve 100% defect-free products, which even tiny defects on their surfaces can significantly impact the performance, accuracy, and longevity of the camshaft. Therefore, precise inspection of camshaft defects in HRS becomes a necessary means of quality control, that reliable inspection methods still cannot meet the requirements. To address these issues, based on a multitask learning (MTL) framework, a lightweight fine-grained perceptual defect inspection network is proposed that is able to optimize both the saliency detection and classification tasks. The network uses a reparameterization technique and a lightweight bottleneck layer that can encode and fuse multiscale features efficiently, while a fine-grained feature-aware module combined with spatial attention and channel attention mechanisms effectively improved the classification accuracy. Based on the experimental validation, the proposed method can achieve a high Mathews correlation coefficient (MCC) of 0.951 and real-time speed of 190 fps on a single GPU, meanwhile demonstrate excellent performance and interpretability in HRS defect inspection and classification.

This paper takes industrial robot disordered object grasping as its research background and proposes an object recognition method based on binocular laser three-dimensional (3D) scanning imaging for the characteristics of grasping objects with complex structures, single surface colors, and scattered stacking. First, to address the issue of incomplete data in the 3D reconstruction of stacked scenes caused by self-obscuration and mutual occlusion of objects, this paper proposes a binocular laser scanning 3D imaging method that utilizes a rapid fusion of multi-source point cloud data. Secondly, this paper reports improved point cloud matching of a pose estimation algorithm based on the point pair feature (PPF) to achieve accurate pose estimation of scattered stacked objects. In the offline training stage, the hidden point removal (HPR) method is introduced to impose visibility constraints, eliminating interference from occluded points and achieving a more accurate global description of the model. In the online matching stage, farthest point sampling (FPS) is applied to the scene point cloud to improve the uniformity of the sampling results. Through Hough voting, candidate poses are quickly obtained. The rough estimation of the object pose was completed by adjusting the grasping pose screening strategy to remove the mismatched poses. Finally, the iterative closest point (ICP) algorithm is used to optimize the poses and obtain more accurate pose results. The experimental results demonstrate that the object information reconstruction completeness of the proposed method is significantly improved compared with the original binocular vision and monocular laser vision system, and the recognition time and average error are reduced compared to the original PPF algorithm. Furthermore, the matching success rate is 100% in 20 cluttered scenes. In the actual grasping experiments, the success rate of grasping is 93.3%, which verifies the proposed method’s effectiveness for cluttered occlusion.

This paper proposes a novel 6D pose estimation network for robotic grasping tasks in complex industrial environments, aiming to address the accuracy limitations of existing sparse featurebased methods. The proposed approach combines a hybrid attention mechanism with local consistency constraint enhancement to construct a two-stage framework. During the training phase, the network adopts ConvNeXt-Tiny as the backbone and utilizes the designed TEAM module to perform selective feature fusion, generating direction vector fields, segmentation masks, and initial pixel weights. During the testing phase, a distance-vector-based local consistency constraint is introduced to refine the initial pixel weights, thereby generating the final weight field used for pose estimation. Experiments conducted on the LINEMOD and Occlusion LINEMOD datasets, as well as simulated industrial grasping scenarios, demonstrate that the proposed method outperforms existing mainstream approaches in 6D pose estimation. This study provides a robust and effective solution for applications such as unordered robotic grasping in industrial settings.

To achieve an automatic unloading of a reactor during the sherardizing process, it is necessary to calculate the pose and position of the reactors in an industrial environment with various amounts of luminance and floating dust. In this study, the defects of classic image processing methods and deep learning methods used for locating the reactors are first analyzed. Next, an improved You Only Look Once(YOLO) model is employed to find the region of interest of the handling hole and a handling hole corner detection method based on the image morphology and a Hough transform is presented. Finally, the position and pose of the reactors will be obtained by establishing a 3D handling hole model according to the principle of a binocular stereo system. To test the performance of the proposed method, a set of experimental systems was set up and experiments were conducted. The results indicate that the proposed location method is effective and the precision of the position recognition can be controlled to within 4.64 mm and $1.68°$ when the cameras are approximately 5 m away from the reactor, meeting the requirements.

Non-standard workpieces and geometric deviations between CAD models and physical parts often cause offline-generated robot trajectories to fail to meet precision requirements. To address this challenge, we present a 3D vision–guided adaptive trajectory correction system that fuses global CAD priors with local point cloud refinements. The perception stage introduces a region-weighted Point Pair Feature (PPF) method with visibility-aware preprocessing and a two-stage verification strategy, enhancing matching accuracy in task-critical regions while maintaining robustness to occlusion. In the trajectory generation stage, a lightweight local correction module detects and refines deviated trajectory segments through direction-guided point cloud projection and Catmull–Rom interpolation, achieving smooth and precise path adjustments without full replanning. Experimental results on rigid workpieces with induced defects show a 100% pose estimation success rate in the evaluated scenarios, an average 41% reduction in Region-of-Interest Root Mean Square Error (ROI-RMSE) compared to baseline PPF, and sub-millimeter trajectory conformity with millisecond-level latency. The proposed framework maintains compatibility with standard offline programming workflows, offering a practical and scalable solution for high-precision robotic guidance in complex industrial environments.