GridFormer: Grid Foreign Object Detection also Requires Transformer

Under the rapid development of China’s energy industry, the invasion of foreign objects poses a considerable challenge to the operation and maintenance of power transmission channels. Due to the particularity of the power system, there needs to be more open-source datasets for power transmission foreign object detection, which limits the development of this field and needs to be addressed urgently. Additionally, existing object detection models are often too complex to meet the real-time inference requirements of drones and other terminal devices. To solve these problems, this paper proposes a lightweight object detection model named GridFormer based on a hybrid feature extraction network. This model combines the advantages of convolutional neural networks (CNNs) and transformers, aiming to improve object detection accuracy and real-time performance. Experimental results demonstrate that the proposed model achieves an mAP value 96.78 on a power transmission foreign object dataset. On an NVIDIA GPU 3080, the inference speed can reach 68.7 FPS with only minor loss compared to GhostNet. The model achieves an mAP value of 79.04 on the Pascal VOC dataset, further validating its effectiveness. Compared to GhostNet, the proposed model exhibits superior performance in terms of object detection performance. By addressing the issues above, the GridFormer model is expected to support the development of China’s energy industry, improve the efficiency of power transmission operation and maintenance, and promote the development of the object detection field.