Multi-scale feature extraction of insulators on distribution network lines from UAV inspection images based on the Faster R-CNN algorithm

To address the challenge of extracting multi-scale features from distribution line insulators in complex inspection environments—where varying scales and background interference complicate detection—this study proposes a multi-scale feature extraction method for (Unmanned Aerial Vehicle) UAV-captured images. The approach leverages the fast region-based convolutional neural network algorithm. Raw images of distribution line insulators captured by UAVs undergo preprocessing—Gaussian filtering for denoising, histogram equalization for contrast enhancement, and gradient texture smoothing—to generate high-quality inputs. Multiscale feature fusion is achieved through superpixel segmentation and feature pyramid networks. Feature discriminative power is enhanced by integrating the lightweight MobileNetv3 backbone with attention mechanisms. A multi-scale anchor box region proposal network and region of interest alignment module are designed to enhance small object detection accuracy and feature space alignment consistency. Contextual information and local complexity features are integrated to form a multi-scale insulator feature representation combining both detail and semantic information. Experiments demonstrate that this method achieves over 90% in six evaluation metrics, including feature diversity and scale coverage, within the test area. When wind speed during UAV inspections increases to 8 m/s, the feature extraction accuracy remains at 96.12% with an extraction time of 0.68 s. This validates the method’s strong adaptability in dynamic environments, providing reliable technical support for subsequent insulator defect identification and condition assessment.