Few-shot Image Information Mining and Its Application in UAV Early Warning

Artificial intelligence applications in vertical fields face the problem of a small and unbalanced number of annotated samples. This article attempts to solve the small sample problem from three perspectives: data, model, and algorithm, by selecting UAV early warning. For data, conventional image enhancement, Mixup enhancement, random masking, and Partial convolution techniques are used to enhance the paradigm of Few-shot samples. Transfer learning techniques are applied to algorithms. Firstly, an autoencoder is trained using many unlabeled images combined with a masking strategy. Secondly, a classification layer is added at the end of the encoding part of the autoencoder. For the model, a multitasking strategy is adopted, which involves both supervised classification and self-supervised image inpainting of annotated images to improve the accuracy of the supervised classification part of the network after transfer learning. Finally, this article draws on the k-nearest neighbor algorithm and metric-based meta-learning ideas and introduces an anomaly warning mechanism, which is combined with conventional detection and warning to reduce the missed alarm rate in UAV early warning.