In order to address issues such as network bias, false positives, and false negatives in few-shot object detection algorithms, a method for foreign object detection in distribution lines based on few-shot object detection was proposed. Firstly, an attention mechanism was introduced during transfer training to mitigate network bias. Secondly, a method for online hard sample selection was proposed to tackle the problems of few-shot classification, as well as false positives and false negatives for small-sized foreign objects. Thirdly, internal convolution was employed to address traditional convolution issues, thereby improving foreign object detection accuracy. Fourth, a novel anchor box scheme was introduced to address the issue of imprecise localization of small-sized foreign objects. Finally, a distribution line foreign object dataset was constructed. Compared to previous state-of-the-art algorithms, this method demonstrates a 4.4% improvement in detection accuracy on the distribution line foreign object detection dataset, reaching 98.6%, with outstanding performance.