Semantic Segmentation Optimization in Power Systems: Enhancing Human-Like Switching Operations

Abstract

In addressing the digital and intelligent transformation challenges within the traditional desulfurization process of coal secondary utilization, a perception learning model rooted in semantic segmentation networks has been developed. This model, when integrated with realworld operational environments, is shown to confront issues arising from symmetry and multi-scale features inherent to thermal power distribution room contexts. A combination of multi-scale feature fusion and attention mechanisms has facilitated the precise detection of human-like operation knobs on switch operation panels. To cope with extended dynamic scenes, a method relying on the visual bag-of-words has been adopted, wherein local image features are extracted and matched against a visual dictionary, resulting in a refined visual representation. The subsequent selection of consecutive symmetrically similar scene keyframes and the elimination of superfluous data have been observed to augment the efficacy of loop-closure detection. Such enhancements have culminated in improved accuracy in the SLAM (Simultaneous Localization and Mapping) of mobile robots, enabling their autonomous navigation to designated switching operation task locations. Experimental findings underscore the superiority of this optimized model over traditional semantic segmentation networks, with its ability to pinpoint operation knobs on electrical control cabinet panels in distribution rooms. Moreover, before initiating grasping actions under the Eye-in-hand architecture, visual servo grasping maneuvers can be executed, irrespective of the target's appearance angle within the field of view. This optimization offers an insightful foundation for potential integrations into patrol operation mobile robots, marking a feasible and effective stride forward.