Feature Weighting and Fusion for Few-shot Object Detection in Remote Sensing Images

Few-shot object detection (FSOD) trains detectors to recognize new objects using very limited data, making it a challenging task in scenarios with scarce labeled examples. In this paper, a small-sample remote sensing target detection method with feature weighting and fusion is proposed by using the meta-learning method. Firstly, the bottleneck structural feature learning module C3 is embedded in the meta-feature extraction network to increase the network depth and receptive field. Secondly, the path aggregation network is used for meta-feature fusion, which effectively improves the perception ability of the network for multi-scale remote sensing targets. Finally, the lightweight convolutional neural network is used to learn the prototype vector to weight the meta-features. The existing knowledge of the model is used to quickly fine-tune the model to adapt to the detection of new types of targets while lightening the model. The experimental results show that the average detection accuracy of the new class of objects reaches 68% and 33% under the 10-shot on the NWPU VHR-10 and DIOR datasets. The visualization results show that the proposed method has a better effect on the detection of small-sample remote sensing targets.