TL;DR
6thGrid-Net is a highly efficient, lightweight image restoration framework designed to simultaneously handle color restoration and edge preservation for remote sensing images on resource-constrained edge devices. By integrating 3D lookup tables (LUTs) and bilateral grids into a single six-dimensional fusion tensor, it eliminates the artifacts and computational overhead of traditional sequential pipelines. This approach achieves state-of-the-art dehazing and cloud removal performance while running at over 70 FPS on mobile devices.
What Problem Does This Paper Address?
Remote sensing images captured by satellites or drones are frequently degraded by adverse weather conditions like clouds and haze. These atmospheric distortions obscure critical surface details, reduce contrast, and cause severe color shifts, impairing downstream applications like urban planning or agricultural assessment.
Existing restoration methods suffer from two major bottlenecks:
- Heavy Computational Burdens: State-of-the-art models often rely on massive deep learning architectures (like Transformers or Diffusion models) that are too computationally expensive to deploy on edge devices like drones.
- Sequential Pipeline Artifacts: Traditional lightweight pipelines handle detail enhancement and color restoration sequentially. As a result, they suffer from mutual interference—over-sharpening details can distort the original color distribution, while color mapping often attenuates high-frequency edge and texture details.
Furthermore, previous grid-based approaches utilize fixed, isotropic interpolation kernels that uniformly mix pixels across edges, leading to blurred boundaries and ghosting artifacts.
Methods
To bypass these limitations, the authors introduce 6thGrid-Net, a parallel integrated framework that couples global color mapping with local detail enhancement in a single forward pass.
The architecture consists of three core components:
- Bilateral Grid Weights Generation: A lightweight U-Net processes the input image's luminance map to predict a spatial-luma weight grid ($W$). This grid adaptively determines whether a local region should favor global color mapping (in smooth areas) or preserve original high-frequency details (along edges).
- 6D Fusion Tensor Construction: The framework constructs a novel six-dimensional tensor spanning spatial coordinates ($X, Y$), luminance ($L$), and RGB color dimensions ($R, G, B$). This tensor seamlessly blends the color rendition capabilities of a learnable 3D LUT with the local detail preservation of the bilateral grid.
- Manifold-Adaptive High-Dimensional Sampling: Replacing standard trilinear interpolation, this novel mechanism dynamically adjusts the sampling kernel per pixel based on local edge orientation, texture strength, and color similarity. It elongates the kernel along edges to maintain continuity and suppresses it across edges to prevent blur and ghosting.
- Edge-Aware Smoothing & Quantization: An edge-aware total variation (TV) regularization loss suppresses noise and blocky artifacts in homogeneous regions. Finally, dynamic post-training quantization (INT8) compresses the network weights for efficient edge deployment.
Datasets
The framework was evaluated on a combination of public benchmarks:
- SateHaze1K: Contains 1,200 pairs of 640×640 images categorized into thin, moderate, and thick haze concentrations.
- RICE (Remote sensing Image Cloud rEmoving): The RICE-I subset was used, containing 500 pairs of 512×512 cloud-contaminated and cloud-free images.
- Unified Dataset: The authors combined SateHaze1K and RICE into a single 1,700-pair dataset to rigorously test "all-in-one" multi-degradation restoration.
- ViCoW: A historical dataset consisting of 1,896 high-resolution (1280×720) grayscale-to-RGB image pairs, utilized to evaluate the model's generalization capability in color recovery tasks.
Key Insights and Ideas
- Decoupling via Higher Dimensions: By embedding the image restoration problem into a unified 6D space, the network avoids the architectural bottlenecks of sequential pipelines. Flat regions rely on the LUT for consistent color styling, while textured regions retain high-frequency details without being over-smoothed.
- Anisotropic Sampling: Moving away from standard isotropic kernels allows the model to respect the natural low-dimensional manifold of images. Stretching the kernel selectively along edges prevents the mixing of pixels from different objects, eliminating edge blur.
- Quantization without Quality Loss: The authors demonstrate that applying dynamic post-training quantization to the U-Net weights yields a massive reduction in computational cost (FLOPs) while maintaining identical peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) metrics.
Key Results
- State-of-the-Art Fidelity: On the unified remote sensing dataset, 6thGrid-Net achieved a PSNR of 22.52 dB and an SSIM of 0.84, outperforming heavy diffusion models and existing real-time networks like LUTwithGrid and CLUT-Net.
- Edge-Device Ready: Following dynamic quantization, the model reduced its computational complexity from 42.01 G FLOPs to 8.10 G FLOPs, achieving a processing speed of 70.83 FPS on a mainstream mobile device.
- Strong Generalization: When tested on the grayscale colorization task using the ViCoW dataset, the model successfully reconstructed vivid, semantically plausible colors (achieving a PSNR of 28.61 dB), proving its versatility beyond simple dehazing.
Limitations and Future Work
Despite its efficiency, the authors acknowledge that the current six-dimensional tensor construction and manifold-adaptive sampling mechanism still incur a moderate memory footprint when scaling up to ultra-high-resolution (e.g., 4K) imagery.
To address this, future work will explore adaptive grid pruning and learnable downsampling strategies to further minimize storage requirements and computational overhead without sacrificing edge and color accuracy.
Summary updated