Adaptive meta-knowledge transfer network for few-shot object detection in very high resolution remote sensing images

Abstract

Object detection on very high resolution (VHR) remote sensing images is a crucial task that has seen remarkable progress in developing deep learning techniques. However, deep learning-based methods rely heavily on the quality and quantity of labeled data. Although few-shot object detection (FSOD) can mitigate this dependency, existing methods still face challenges, including domain shifts between base and novel classes, misclassification due to class similarities, and limited ability to acquire effective information from a few samples. VHR remote sensing images exacerbate these issues due to their greater intra-class diversity and weaker inter-class separability. To address these issues, we propose a new FSOD network, the Adaptive Meta-Knowledge Transfer Network (AMTN). AMTN adaptively transfers meta-knowledge from the source domain to the target domain by effectively obtaining valid information and keeping a stronger discriminative ability towards objects of similar classes in low-shot scenarios. Specifically, considering that VHR remote sensing images’ higher frequency domain resolution provides richer frequency domain information, we employ a Spatial-Frequency Joint Enhancement (SFJE) model to achieve dual enhancement of query image features by fusing information across the spatial and frequency domains. Moreover, we propose the Adaptive Reweighting (AR) loss to enhance the sensitivity of novel class detection in limited sample scenarios. Finally, we propose Sim-Meta loss for discriminability enhancement of similar classes. Our AMTN has demonstrated effectiveness and stability through multiple rounds of experiments on the largescale DIOR dataset while also achieving rapid transfer of meta-knowledge.

Keywords

• Few-shot learning
• Object detection
• Convolutional neural network
• Deep learning
• Remote sensing