Ensuring the reliability, stability, and effectiveness of the power transmission system and early detection of potential risks in power transmission lines are crucial. Deep learning-based detection methods offer more efficient detection of abnormal targets on power transmission lines. However, challenges such as complex backgrounds, large-scale variations and numerous small objects exist in power transmission lines insulator. To address these challenges, we proposes INet. To tackle the issue of complex backgrounds, the large selective kernel attention is introduced, which conducts multi-scale convolution operations on input feature maps and combines with an adaptive attention mechanism to focus the model’s attention on object areas. To address the problem of large- scale variations in objects, the bidirectional feature fusion module is introduced, which effectively integrates multi-scale features through a cross-scale information interaction mechanism, thereby enhancing the model’s detection capabilities for objects of different scales. Finally, to tackle the challenge of difficulty in regression of small objects, the Inner-IoU loss is introduced, which penalizes deviations between objects and bounding boxes, promoting accurate detection of small objects by the model. INet, incorporating these key components, achieves promising results on the Insulator dataset, providing reliable technical support for the safe operation of power transmission systems.