Graph convolution detection method of transmission line fitting based on orientation reasoning

To address object occlusion resulting from the density of multiple fittings in transmission lines, a novel graph convolution detection method based on orientation reasoning is proposed. Firstly, the spatial relationship between different categories of fittings was analyzed through a UAV inspection shooting standard. The relative category orientation concept was introduced to express the orientation relationship between the structures of fittings in a data-driven manner. To incorporate spatial orientation information into the deep learning model, the visual features of ROI (region of interest) results were treated as nodes of a spatial connection graph. The regional orientation adjacency matrix obtained by adaptive learning was integrated as relations of the graph. Subsequently, a graph convolutional network was employed to establish the orientation reasoning model. Experimental results were conducted on a dataset(14 categories of fittings). The proposed model outperformed other advanced object detection models in terms of overall detection effect. Compared to the baseline model, the proposed model increased mean average precision by 6.3%. Ablation experiments further confirmed that each module contributes to the improved detection effect. This proposed approach combines advantages of orientation reasoning and graph convolutional networks to enhance average detection accuracy and effectively overcome object occlusion issue.