Research on a high-precision extraction method of industrial cuboid

The precise extraction of the cuboid takes on a critical significance in the industry. The research on this cuboid extraction has not attracted critical attention for its particularity and complexity. A coarse-to-fine industrial cuboid extraction method is proposed in this study, which exhibits high precision, reliability, and robustness. The method is implemented in two steps. At the coarse extraction stage, a lightweight parallel connection network (PaRNet) serves as the backbone network, and industrial cuboid detection is implemented based on cross-layer feature fusion cross-layer connections. Subsequently, the cuboid boundaries will be well retrieved based on the Screening area Line-segment Recombination (SLR) in fine extraction. The proposed method overcomes the difficulties in multi-cuboid extraction and solves the problem of edge discontinuity, boundary duplication, damage interference, and vertex selection in cuboid edge extraction. We tested the proposed method on a low signal-to-noise ratio industrial cuboid image dataset, and the results show that the average precision (AP) for non-overlapping cuboid object detection can reach 96.28%, the missed detection rate (MDR) is only 1.36%, and the false alarm rate (FAR) is lower than 0.01%. In addition, SLR performs well in terms of precision, recall, F1-measure, and reliability for cuboids without obvious interference, all of which can reach test results of more than 95%, and the extraction time for a single cuboid is only 0.27s. Furthermore, SLR has an optimal error of less than 0.1 mm and an average measurement error of less than 0.5 mm for cuboid size measurement, demonstrating high accuracy for edge extraction of cuboids.