Channel Attention and Partial Convolution for Normalizing Flows in Power Line Defect Detection

In power line defect detection, due to the lack of annotated defective component images and the stacking of convolutional modules in the model, defect detection faces drawbacks such as low accuracy, model complexity, and computational inefficiency. To address this issue, a novel approach is presented where partial convolution and Squeeze-and-Excitation Networks are integrated into PCSENet. Subsequently, cross-scale normalizing flow (CS-Flow) is enhanced with PCSENet to achieve effective defect detection at the image level. Specifically, the fully convolutional network in the original method is replaced with PCSENet. This substitution reduces the network's Flops and parameters while improving the model's feature extraction capability. We conducted experimental analysis utilizing five common defects from the InsPLAD dataset. Experiments show that our method achieved a higher AUROC than in 4 categories. The average AUROC across all categories also surpassed the original method. Also, our model has approximately 52.74% fewer parameters and 52.19% fewer FLOPs than the original one. These results demonstrate the effectiveness and superiority of our method for power line fault detection.