An object detection model for power equipment based on SVGA-Net

Abstract

Convolutional neural networks （CNNs） struggle to efficiently capture contextual information of power equipment such as arresters and GIS inlet casings due to their limited receptive fields， thereby affecting detection performance. To address this issue， the paper introduces a Transformer-based voxel-graph attention network. Local attention and dilated attention mechanisms are proposed to respectively capture short-range and long-range feature correlations within image volume pixels， effectively expanding the attention scope while keeping computational costs unchanged. Additionally， submanifold voxel modules and sparse voxel modules are designed to extract feature information from non-empty voxel positions and empty voxel positions， respectively. Finally， through comparative analysis with mainstream models on the general datasets Waymo and KITTI， as well as on an image dataset from a transmission and transformation area in Yunnan Province， the superior performance of the proposed model in detecting power equipment is demonstrated.

Keywords

• attention network
• object detection
• geometric manifold
• volume pixel