Abstract: In order to enable the model to learn new categories using a small number of samples, the cross-image dual-similarity few-shot object detection method based on meta-learning is proposed. We proposed the feature self-enhancement module to enrich the representation capabilities of supporting features. In order to assign support information to query features, the cross image similarity enhancement module between features was designed based on cross-attention mechanism. Cosine similarity was introduced in the classification header to calculate the distance metric between the query vector and the support vector for correcting the classification prediction score of the network. The test results on the Pascal VOC dataset show that the improved algorithm can effectively improve the target detection ability for new categories, and improve the average mean accuracy mAP₀.₅ with an IoU threshold of 0.5 increased by 4.1 percentage points.