Efficient Meta-Learning Enabled Lightweight Multiscale Few-Shot Object Detection in Remote Sensing Images

TL;DR

This work tackles few-shot object detection in remote sensing images by introducing LMFSODet, a lightweight, one-stage detector based on YOLOv7 augmented with meta-learning components. The method integrates meta-sampling, a meta-cross loss, and a meta-cross category decision mechanism, coupled with an Enhanced Class-Weighted Evaluation Standard to balance base and novel class performance. Evaluations on the DIOR and NWPU VHR-10.v2 datasets show LMFSODet achieving strong detection accuracy and speed, particularly in higher-shot regimes, while maintaining low parameter counts. The approach demonstrates practical potential for scalable FSOD in large-scale RSIs, with future work aimed at deeper exploration of meta-sample information and further lightweight meta-feature extraction.

Abstract

Presently, the task of few-shot object detection (FSOD) in remote sensing images (RSIs) has become a focal point of attention. Numerous few-shot detectors, particularly those based on two-stage detectors, face challenges when dealing with the multiscale complexities inherent in RSIs. Moreover, these detectors present impractical characteristics in real-world applications, mainly due to their unwieldy model parameters when handling large amount of data. In contrast, we recognize the advantages of one-stage detectors, including high detection speed and a global receptive field. Consequently, we choose the YOLOv7 one-stage detector as a baseline and subject it to a novel meta-learning training framework. This transformation allows the detector to adeptly address FSOD tasks while capitalizing on its inherent advantage of lightweight. Additionally, we thoroughly investigate the samples generated by the meta-learning strategy and introduce a novel meta-sampling approach to retain samples produced by our designed meta-detection head. Coupled with our devised meta-cross loss, we deliberately utilize "negative samples" that are often overlooked to extract valuable knowledge from them. This approach serves to enhance detection accuracy and efficiently refine the overall meta-learning strategy. To validate the effectiveness of our proposed detector, we conducted performance comparisons with current state-of-the-art detectors using the DIOR and NWPU VHR-10.v2 datasets, yielding satisfactory results.

Figures (15)

• Figure 1: The comparative results of the accuracy of our proposed detector LMFSODet, against state-of-the-art detectors under the 10-shot scenario on the DIOR dataset.
• Figure 2: Dataset Partitioning Diagram. Please note that the actual number of instances for each class may vary; the illustration in the figure is for visual representation purposes only, and the sizes are not to scale.
• Figure 3: Proposed meta-learning-based FSOD framework. It comprises the meta-feature extraction module, support set feature extraction module, channel-wise multiplication, meta-detection head and meta-cross loss. In contrast to mainstream meta-learning design approaches, we place the design emphasis on the post-processing of output features, specifically the meta-sampling technique and loss computation, which fully leverages the meta-learning design philosophy, pushing the advantages of meta-learning solutions for FSOD.
• Figure 4: Comparison of samples handling patterns among Meta-RCNN meta-rcnn, FRW FRW, and the network we propose (LMFSODet).
• Figure 5: The illustration of the loss computation process. The loss comprises four components, namely the coordinate loss, object confidence loss, classification loss and meta-cross loss, each governed by the corresponding principles illustrated in the rightmost diagram. Leveraging the benefits of meta-sampling, we conduct loss computation utilizing the informative samples, leading to a significant enhancement in sample utilization efficiency.