MSP-Net: A multi-scale strip pooling-based network for fine segmentation of anti-vibration hammer defects

Abstract Defects in anti-vibration hammers on overhead transmission lines pose a severe threat to the operational safety of power grids. Existing multi-defect segmentation models frequently suffer from feature discontinuity when processing extremely slender steel strands, along with blurred edges of microscopic rust spots and extreme class imbalance. To address these issues, a multi scale strip pooling network (MSP-Net) is proposed in this paper. Specifically, the stem layer at the front end of the ResNet34 encoder is reconstructed using a combination of small multi-scale convolutions to enhance the capability of capturing fine corrosion edges and texture details. Furthermore, the concept of strip pooling is adapted to multi-scale scenarios to construct the multi-scale strip pooling module (MSPM), which effectively mitigates the feature discontinuity problem commonly observed in industrial defect imaging. Subsequently, a dual-weighted Focal Dice joint loss function is formulated to reinforce supervision from the perspectives of pixel classification difficulty and region proportion, effectively improving the segmentation accuracy of minute corrosion defects against complex backgrounds. Extensive experiments on a custom dataset demonstrate that MSP-Net achieves an mIoU improvement of 7.3 percentage points over the baseline model, operating at an inference speed of 130 frames per second with a computational complexity of 41.8 GFLOPs. These results demonstrate that MSP-Net provides an effective and efficient solution for automated transmission line inspection.