In the field of industrial automation, the accurate detection and instance segmentation of flexible cables constitute a crucial prerequisite. This process is essential for achieving stable and precise grasping operations. Therefore, this study aims to address this need. The primary goal is to significantly improve the instance extraction precision of flexible cables by improving existing methods. This enhancement provides a reliable visual perception foundation for robotic manipulation.
To enhance the mask segmentation accuracy of flexible cables in complex environments, this study constructs the HCD-Net (High-resolution channel-spatial detail network) model. HCD-Net uses HRNet (High-resolution network) as the backbone network. The model introduces the channel-space fusion attention (CSFA) mechanism to strengthen the binary segmentation capability for multiple cables. Furthermore, the authors propose a novel approach involving multiscale key point detection and a scale-adaptive graph representation method. This pipeline achieves the effective segmentation and extraction of cable instances with significant thickness variations through graph modeling, intersection handling and linear optimization.
The proposed method achieves a mean Intersection over Union (mIoU) of 94.31% on the self-built data set, which significantly outperforms other mainstream methods. In addition, the proposed method accurately determines the spatial relationships at cable crossings, thereby significantly enhancing the completeness and accuracy of the extraction results.
In real-world industrial scenarios, cable widths often exhibit significant variations. Existing instance extraction methods struggle to accurately determine cable ownership in regions with abrupt width changes, frequently leading to robotic gripper failures. Through multiple rounds of experimental validation, this approach demonstrates precise identification and segmentation of diverse cables within complex backgrounds, providing robust technical support for stable and reliable robotic manipulation.
