.png){kind=icon}
.png){kind=icon}
Return
RingMo-Qlite: A Fully Quantized Lightweight Network for Efficient Remote Sensing Image Interpretation
.png){kind=icon}
PRE
AI
DOI：10.1109/TGRS.2026.3656351.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With the development of remote sensing satellite technology, on-orbit interpretation of remote sensing images has become a new trend. Hybrid network architectures, such as those combining convolutional neural networks (CNNs) and transformers, have shown great potential in remote sensing image interpretation due to their ability to capture local features and long-range dependencies. However, these architectures face limitations in real-time on-orbit processing due to their large parameter sizes and high computational complexity. To address this, we propose RingMo-Qlite, a fully quantized remote sensing model designed for efficient processing. All weights, activations and nonlinear layers are quantized to low-bit integers, with inference relying solely on integer and shift operations. We introduce a training-free token mixer to replace the multihead self-attention mechanism in the transformer branch, effectively reducing precision loss during quantization. For the CNN branch’s high-frequency feature extraction, we employ a reparameterization strategy that simplifies the multibranch structure used during training to a single-branch structure during inference, significantly reducing computational load. In addition, we design a novel quantization strategy that includes a peak-aware quantizer and an outlier-robust observer, tailored to the complex distribution of remote sensing images to minimize precision loss. Experiments on various remote sensing scene classification datasets demonstrate that RingMo-Qlite achieves higher accuracy than existing lightweight methods, with near-lossless performance under INT8 quantization (less than 1% accuracy loss) and comparable performance under INT4 quantization.
Keywords:
Convolutional neural network (CNN)-transformer hybrid framework
model quantization
posttraining quantization (PTQ)
remote sensing frequency domain features
reparameterization
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W