Abstract:The overwhelming majority of models for remote-sensing image (RSI) scene classification generally require the weights pretrained on natural images for initialization before formal training. However, differences in imaging mechanisms lead to huge discrepancies between natural images and RSIs, and the strong visual representation learned from massive natural images limits the performance of models when inferencing RSIs. To address this issue, the well-established self-supervised contrastive learning paradigm in the natural image field is introduced to the RSI field. We propose a contrastive learning method based on multiscale hard features (MHCL), which aims to use finite RSIs to learn sufficient visual representations in an unsupervised contrastive manner, thus providing a powerful upstream pretrained model for fine-tuning downstream scene classification task. Multilevel features extracted by intermediate layers of each encoder's backbone are first gathered and then a hard features transformation method (HFT) is proposed to create hard positive features and diverse queues that save hard negatives, thereby enriching the finite scene information in small-scale RSIs. Furthermore, we redesign the multiscale hard features joint contrastive loss to boost the model to explore sufficient invariant representations by additionally pulling hard positive pairs closer and pushing hard negative pairs farther away in the embedding space. Extensive experiments demonstrate that the upstream pretraining model generated by MHCL achieves competitive transferred performance on three popular scene classification datasets, outperforming the traditional model pretrained on ImageNet and models pretrained by other state-of-the-art contrastive learning methods. Our code will be released at: https://github.com/benesakitam/MHCL.

Multiscale Object Contrastive Learning-Derived Few-Shot Object Detection in VHR Imagery

FSOD4RSI: Few-Shot Object Detection for Remote Sensing Images via Features Aggregation and Scale Attention

Mutually Reinforcing Structure with Proposal Contrastive Consistency for Few-Shot Object Detection.

Few-shot Oriented Object Detection with Memorable Contrastive Learning in Remote Sensing Images

Few-shot Weakly-Supervised Object Detection via Directional Statistics

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

MPF-Net: multi-projection filtering network for few-shot object detection

Few-Shot Object Detection Via Context-Aware Aggregation for Remote Sensing Images.

Few-Shot Object Detection with Self-Adaptive Global Similarity and Two-Way Foreground Stimulator in Remote Sensing Images

Few-Shot Object Detection with Memory Contrastive Proposal Based on Semantic Priors

Few-shot Object Detection with Feature Attention Highlight Module in Remote Sensing Images

Few-Shot Object Detection for Remote Sensing Imagery Using Segmentation Assistance and Triplet Head

Semi-Supervised Contrastive Learning for Few-Shot Segmentation of Remote Sensing Images

MM-FSOD: Meta and metric integrated few-shot object detection

Information Extraction Enhancement for Few-Shot Object Detection in Remote Sensing Images

Few-Shot Object Detection of Remote Sensing Images via Two-Stage Fine-Tuning

Contrastive Learning Based on Multiscale Hard Features for Remote-Sensing Image Scene Classification.

Few-Shot MS and PAN Joint Classification with Improved Cross-Source Contrastive Learning

Dual Contrastive Network for Few-Shot Remote Sensing Image Scene Classification

Multispectral Object Detection Based on Multilevel Feature Fusion and Dual Feature Modulation

Few-Shot Object Detection With Multilevel Information Interaction for Optical Remote Sensing Images