Abstract
The Global Information Compensation Network (GICN) addresses the persistent trade-off between local feature extraction and global context modeling in image denoising. While standard CNNs excel at local textures, they often lack the long-range dependencies captured by computationally expensive Transformers. GICN bridges this gap by integrating frequency-domain analysis via Fourier transforms and cross-scale feature aggregation. Experimental results demonstrate that GICN achieves state-of-the-art performance on synthetic and real-world datasets while maintaining a significantly lower parameter count and faster inference speed than heavy-duty Transformer models.
Introduction
Image denoising is a fundamental task in computer vision, yet it faces a persistent "local vs. global" trade-off. Standard Convolutional Neural Networks (CNNs) are highly effective at extracting local features such as textures and edges due to their inductive bias and small receptive fields. However, they often struggle to capture global context—the relationships between distant pixels—which is essential for distinguishing complex noise from actual image structures.
While Transformers have emerged as a solution for modeling long-range dependencies, they come with high computational costs and massive parameter counts, making them impractical for resource-constrained environments. The authors of GICN propose a network that "compensates" for this missing global information within a lightweight CNN framework, ensuring high efficiency without sacrificing the ability to perceive the entire image structure.
Proposed Method
GICN utilizes an encoder-decoder (U-Net style) architecture enhanced by four novel components designed to capture both local and global features efficiently.
- Global Feature Mining Block (GFMB): Located at the shallow part of the network, this block uses a dual-branch approach. One branch handles spatial information using non-local blocks, while the other utilizes the Fourier Transform to analyze the image in the frequency domain. Since a single point in the frequency domain represents global patterns in the spatial domain, this allows the network to capture global structure early in the pipeline.
- Residual Dynamic Convolution Block (RDCB): Instead of static filters, GICN employs dynamic convolutions that adapt their weights based on the input. This provides a flexible receptive field, allowing the model to adjust its focus depending on local noise characteristics.
- Cross-Scale Feature Aggregation Block (CSFAB): Positioned between the encoder and decoder, this block fuses features from multiple resolutions. This ensures that fine-grained details from high-resolution layers and semantic context from low-resolution layers are combined effectively.
- Skip Connection Attention (SK-Att): Standard skip connections can sometimes pass noisy information directly to the decoder. GICN introduces a channel attention mechanism in these connections to dynamically weight and filter the information flow, prioritizing the most useful features for reconstruction.
Experimental Evaluation
The researchers evaluated GICN on a comprehensive suite of benchmarks. For synthetic Gaussian noise, the model was tested on BSD68, Kodak24, McMaster, and CBSD68. Real-world noise performance was assessed using the SIDD (Smartphone Image Denoising Dataset), Nam, and PolyU datasets. Training was conducted on a composite dataset of 4,200 images for synthetic scenarios, while SIDD and RENOIR were used for real-world noise training.
Results and Discussion
GICN consistently outperformed mainstream methods across multiple benchmarks. On the McMaster dataset (noise level $\sigma=35$), GICN achieved a PSNR of 31.21 dB, outperforming advanced models like DRANet (31.08 dB) and MWDCNN (31.05 dB). In real-world tests on the Nam dataset, GICN reached a PSNR of 40.25 dB, significantly higher than DnCNN (34.95 dB).
Beyond quantitative metrics, GICN strikes an impressive balance between performance and efficiency. It uses only 2.98M parameters, compared to 20.63M for Uformer-S, yet delivers comparable or better denoising results. Its inference time on a 256x256 image is just 0.03 seconds, making it roughly 29x faster than Uformer. These results highlight that frequency-domain operations can serve as an efficient shortcut to global context, providing "Transformer-like" awareness with a fraction of the overhead.
Conclusion
The Global Information Compensation Network demonstrates that targeted global information recovery can significantly enhance CNN-based denoising. By utilizing the Fourier Transform and cross-scale aggregation, GICN achieves state-of-the-art results with high efficiency. Future work may explore the generalization of this "global compensation" strategy to other low-level vision tasks, such as deblurring or super-resolution, and further optimize frequency-domain operations for deployment on edge hardware.
Summary updated