An adaptive dual-weighted feature network for insulator detection in transmission lines

In the field of electrical power applications, high-voltage insulators necessitate routine inspection to assure the security and stability of the whole electric power system operation. Accurately positioning the insulator is extremely crucial for proceeding to the insulator defect detection. However, during UAV electrical line inspection, the presence of the electric power line magnetic field engenders a reduction in the pixel representation of the insulator within the image data, thereby diminishing the accuracy of insulator detection. In response to the prevailing issues, we present the creation of the adaptive dual-weighted feature network in this paper. Simultaneously, we create an insulator dataset to substantiate the effectiveness of enhanced model in detecting small insulators. Firstly, the integration of context fusion network is employed to capture comprehensive contextual features for each effective feature map. In addition, a cross-scale residual perception network is incorporated into the neck prior to three concatenation modules, facilitating the collection of diverse information across levels. Finally, a Dual-Weighted Feature Fusion module is designed to replace the conventional concatenation pattern within the neck, thus achieving a more precise representation of object features. Experiments are conducted on the insulator dataset, the RSOD dataset and the NWPU VHR-10 dataset to evaluate the designed model, resulting in mAP values that were 3.92%, 1.55% and 2.39% higher than the YOLOv7, respectively.