This paper proposes a method based on cross-domain learning to address the problem of small sample sizes in one-shot object detection. The proposed method begins with the aim of data enhancement and progresses with the addition of datasets in other domains as auxiliaries to enhance the network learning capabilities, simultaneously a cross-domain learning algorithm based on image and instance scales is proposed to solve the problem of differences between domains. A domain classifier model is added to the input image features and candidate features of the detection network to enhance the background of the network to cross-domain data and the target domain adaptability. Experiments for two different cross-domain scenarios are conducted, the PASCAL VOC dataset is compared with current mainstream one-shot object detection algorithms, and it presents an improvement of 2.8 percentage points on the current best algorithm. This proves that the proposed method can effectively improve the performance of one-shot object detection algorithm.