Abstract
Due to the tiny size and dense connection of the objects in the transmission line components, the occlusion between the objects is severe, resulting in the problems of missed and false detection of the transmission line components. We propose a tiny object detection method for transmission lines based on dynamic-NMS to solve the problems. First, we used the Gaussian kernel function method to automatically generate the density map labels from the object detection labels. Next, the CSRNet density estimation network model was trained to predict the transmission line objects' density adaptively. Finally, the predicted density values were mapped to the dynamic NMS threshold to post-process the object detection model and accurately identify unevenly distributed objects through different NMS thresholds. In the experimental stage, we have implemented the dynamic NMS method on three object detection models. Compared with the original model, the test results have been improved, with a maximum increase of 6.2%. Among them, we achieved better performance in detecting dense tiny objects.
Funding Information
  • State Grid Hebei Electric Power Co., Ltd (2020AAA0107500)

This publication has 21 references indexed in Scilit: