Abstract:To address the challenges of low detection accuracy for multi-type defects in transmission lines under complex scenarios and high miss detection rates with single-modality methods, this study proposes a lightweight bimodal object detection framework. The framework employs a three-stage processing pipeline: firstly, EfficientNetV2-S is adopted to perform accurate pre-classification on the input images, then different defect modalities are input into corresponding detection networks, and finally, a decision-level fusion module is activated only when paired bimodal images exist in the same scenario. Additionally, confidence weighting and positional matching of the detection results of both branches are conducted to achieve precise complementarity between visible-light structural information and infrared thermal information. Experimental results show that the proposed framework yields excellent performance on the self-constructed bimodal dataset and multiple public datasets. The infrared branch reaches an F1-score of 0.985 for connection-point overheating defects, while the visible-light branch yields F1-scores of 0.920 and 0.964 for insulator structural defects and foreign object detection, respectively. The overall performance significantly outperforms mainstream CNN-based and YOLO-based methods. The framework supports efficient independent operation in single-modality mode and substantially reduces the missed detection rates for composite defects and complex illumination conditions in bimodal mode.