Initialization Parameters Of Our Method.

The defects of insulators exhibit characteristics such as complex backgrounds, multi-scale variations, and small object sizes. Therefore, accurately focusing on these defects in dynamic and complex natural environments while maintaining inference speed remains a pressing challenge. To address this issue, this paper proposes an innovative insulator defect detection network, TLINet. First, a Multi-Branch Partially Transformer Block (MBPTB) is designed to enhance the backbone’s capability in capturing global features. Next, a Dynamic Downsampling Module (DyDown) is introduced to mitigate the issue of small-scale defect information blurring. Furthermore, considering the multi-scale variations of insulator defects, this paper proposes a Context-Guided Feature Fusion Network (CGFFN). This module enables fine-grained fusion of features at different scales, allowing the model to generate adaptive responses to defects of various sizes. Compared to the baseline model, the proposed method improves mAP50 by 5.3% on our self-constructed Insulator-DET dataset. On CPLID-D and CPLID-N, it achieves mAP50-95 improvements of 7.9% and 12.1%, respectively. Additionally, to verify the robustness of the proposed algorithm, TLINet is evaluated on the VOC07 + 12 dataset. Compared to the baseline model, TLINet improves mAP50 by 0.4% while reducing the number of parameters by 1/6. These results demonstrate the effectiveness of TLINet in addressing the complexities of insulator defect detection in power transmission lines. The code is available at https://github.com/mazilishang/TLINet.