Abstract
In this paper, we proposed a guidance correlation pixel sampling and aggregation image denoising for maritime Unmanned Aerial Vehicles (UAVs) photography, providing a reliable data base for maritime reconnaissance work. The overall step is mainly composed of two parts: pixel sampling and pixel aggregation. To improve the content correlation of sampled pixels, we propose a guided sampling scheme based on the basic estimated map and extend this algorithm to the restoration of maritime UAV images. Finally, a UAV image denoising system is shown. Our experimental results show that the proposed algorithm can effectively remove noise and achieves 32.98 dB and 31.83 dB in Set12 and BSD68 datasets with less image distortion. In the actual scene, the PSNR of our denoising algorithm has reached 35.33 dB, meeting the basic needs of practical vision and follow-up research.
Introduction
Maritime Unmanned Aerial Vehicles (UAVs) are essential for reconnaissance and search-and-rescue, but their cameras often struggle with environmental "degraders" like fog, rain, and low-light conditions. These factors introduce significant noise, which cripples downstream computer vision tasks like ship detection and identification.
Standard denoising techniques often fall into two traps. First, deep learning models like DnCNN use fixed convolution kernels that can over-smooth and blur critical fine details, such as ship structural textures. Second, advanced "pixel aggregation" methods that try to find similar pixels across an image often get "lost" when noise is high, picking unrelated pixels and causing image distortion.
Proposed Methodology: MGPAN
To address these challenges, we developed the Maritime Guidance Pixel Aggregation Network (MGPAN). The framework evolves from the standard Pixel Aggregation Network (PAN) through three key innovations.
1. Guided Sampling and Coordinate Offsets
The core of MGPAN is the use of an Offset Estimated Network (OEN) and a Weight Estimated Network (WEN). Instead of relying on a rigid grid, the OEN predicts offsets to locate the most relevant pixels within a neighborhood. To ensure accuracy in high-noise environments, we introduce Guided Sampling. A preliminary "basic estimated map" (a rough denoised version of the image) serves as a guide, allowing the network to calculate offsets more reliably than it could from the original noisy data.
2. Pixel Aggregation Process
The sampling process combines a standard rigid grid with these learned normalized offsets. This allows the model to "reach out" and aggregate pixels that follow the actual edges and textures of the maritime scene, rather than being confined to a square kernel. The WEN then determines the appropriate weighting for each sampled pixel to produce the final denoised result.
3. Color Domain Extension
Recognizing that human perception is most sensitive to luminance, MGPAN operates in the YCbCr color space. Offsets are estimated using only the Y (luminance) channel and then applied to aggregate pixels across all color channels. This ensures the structural "skeleton" of the image is preserved even when color information is heavily degraded.
Experimental Results
The MGPAN algorithm was evaluated using standard benchmarks (Set12, BSD68, DIV2K) and a custom Maritime UAV Dataset consisting of 677 high-resolution photographs.
- Quantitative Performance: MGPAN achieved PSNR values of 32.98 dB on Set12 and 31.83 dB on BSD68. In real-world maritime testing, it reached 35.33 dB, demonstrating its practical utility for reconnaissance.
- Resilience to High Noise: The model significantly outperformed traditional methods like BM3D and DnCNN, particularly at noise levels where $\sigma > 25$. While other models produced "muddy" results, MGPAN maintained sharp edges and recognizable textures on ship hulls and sea surfaces.
Conclusion and Future Work
MGPAN provides a robust framework for cleaning noisy maritime photography. By leveraging guided sampling and luminance-prioritized aggregation, it avoids the pitfalls of over-smoothing and distortion common in earlier deep learning models. Future research will focus on model distillation to reduce parameter count for real-time deployment on low-power UAV hardware and integrating the denoising engine into end-to-end maritime object detection systems.
Summary updated