Insulator localization and defect detection in multispectral images based on improved YOLOv11s

Insulator inspection is critical for high-voltage power line maintenance. However, traditional methods for detecting insulators and their defects under multiple spectral images suffer from issues such as insufficient accuracy, large number of model parameter, and an elevated incidence of false positives and missed detections. To overcome these limitations, this paper develop an enhanced YOLOv11s architecture for insulator defect identification. Firstly, a lightweight AKConv (Alterable Kernel Convolution) module is introduced into the Backbone, leveraging its adaptive multi-scale convolution characteristics to efficiently extract features of insulators and their defects across multiple spectra while reducing the number of parameters, thereby enhancing defect detection capabilities. Secondly, a CA (Coordinate Attention) mechanism is incorporated into the Neck to improve the spatial localization accuracy of insulators and their defects. Finally, the optimization framework utilizes Inner-IoU (Internal Intersection over Union) loss to enhance bounding box regression accuracy by refining overlap region estimation between predictions and ground truth, consequently improving localization precision while minimizing both false alarms and detection omissions. Experimental results demonstrate that the improved model achieves an mAP@0.5 of 98.3%, a detection speed of 385 frames per second, and a memory footprint of only 8.1MB. The enhanced detection capability and operational efficiency of the model offer reliable technical support for transmission system safety maintenance.