Researchers from the Changchun Institute of Optics, Fine Mechanics and Physics have developed a new deep-learning method for blind Gaussian image denoising, a task that aims to recover clear images when the noise level is unknown in advance. By introducing a progressive way to separate image features at different scales and a directional strategy for preserving edges and textures, the team designed a denoising network that removes noise while better retaining fine structures. The method could support a wide range of imaging applications, from photography and remote sensing to medical imaging and machine vision. The research results were published in Knowledge-Based Systems.

Image denoising is one of the most fundamental problems in computer vision. In practice, images captured by cameras and sensors are often degraded by noise during acquisition, transmission, or storage. This problem is especially serious in low-light conditions, high-sensitivity imaging, and complex real-world environments. If the noise is not effectively removed, it can affect not only image quality itself but also later tasks such as target recognition, image segmentation, scene understanding, and medical diagnosis. Among different noise models, Gaussian noise is widely studied because it approximates many kinds of sensor noise and serves as a standard benchmark for evaluating image restoration methods.

Over the past decade, deep learning has greatly improved image denoising. Convolutional neural networks and U-Net-like structures have made it possible to suppress noise more effectively than many traditional methods. However, blind Gaussian denoising remains difficult because a network must remove noise without being told how strong the noise is. At the same time, real images contain structures at many different scales, from broad smooth regions to tiny textures and sharp edges. Existing methods often mix these features together during processing. As a result, they can fall into a familiar trade-off: removing noise aggressively but blurring image details, or preserving details but leaving visible residual noise and artifacts.

To address this problem, the research team proposed a new network called PSDANet, short for Progressive Feature Separation and Directional Feature Aggregation Network. The model is built on a U-Net-like encoder–decoder framework, but its core innovation lies in two specially designed modules that help the network handle image structures more selectively.

The first module, called the Progressive Feature Separation Module, is designed to deal with the mixing of image features across scales. Instead of treating all features in the same way, it uses dilated convolutions with different receptive fields and a progressive differencing strategy to separate coarse structures, medium-scale information, and fine details into different feature bands. In simple terms, the network learns to distinguish broad image content from delicate textures and from noise-like fluctuations, allowing denoising to be performed in a more targeted way.

The second module, called the Directional Feature Aggregation Module, focuses on the fact that many important image structures are directional. Edges, contours, and repeated textures often have strong horizontal, vertical, or oriented characteristics, but ordinary convolution kernels do not explicitly model these directional cues. To solve this, the team introduced asymmetric convolutions in multiple branches so that the network could learn and combine orientation-sensitive features more effectively. This design improves the preservation of edges and intricate textures that are easily damaged during denoising.

The two modules are integrated into a unified denoising architecture, enabling the network to process multi-scale and directional information at the same time. In effect, the method is not only trying to remove noise from an image; it is also learning how to distinguish between what is likely to be meaningful structure and what is likely to be random corruption. This makes the model especially suitable for blind denoising, where the degradation level is not known beforehand.

To evaluate the method, the researchers carried out experiments on a series of public benchmark datasets for both grayscale and color image denoising, as well as real-world noisy image datasets. On grayscale denoising tasks, PSDANet achieved strong performance across different noise levels. For example, on the Set12 dataset it reached average PSNR values of 33.16 dB, 30.86 dB, and 27.81 dB at noise levels of 15, 25, and 50, respectively. On the BSD68 dataset, it also delivered consistently competitive results, including 29.45 dB at a noise level of 25 and 26.56 dB at a noise level of 50.

The model also performed well on color image denoising benchmarks. On the CBSD68 dataset, PSDANet achieved 34.28 dB, 31.67 dB, and 28.48 dB under the same three noise settings. On Kodak24, it reached 35.17 dB at a noise level of 15 and 32.74 dB at a noise level of 25, while on McMaster it achieved 35.31 dB and 33.06 dB at noise levels of 15 and 25. Beyond PSNR, the researchers also reported strong results in structural similarity and perceptual quality, indicating that the method does not simply smooth away noise, but can preserve visually important details more effectively than many competing methods.

The team further tested the model on real-world noisy image datasets including SIDD and DND. On SIDD, PSDANet achieved a PSNR of 39.64 dB and an SSIM of 0.958. On DND, it reached 39.72 dB in PSNR. Visual comparisons showed that the method could preserve letters, edges, and texture details more clearly than several existing denoisers, especially in challenging regions where other models tended to blur structures or introduce artifacts.

Beyond the performance gains themselves, the work highlights an important direction in image restoration research. Rather than treating denoising as a purely uniform filtering problem, the new method emphasizes structured understanding of image content: separating features by scale, identifying them by orientation, and then restoring them in a more informed way. This idea could be valuable not only for denoising, but also for other low-level vision tasks such as deblurring, super-resolution, and image enhancement.

The researchers believe the method could provide useful support for many real-world imaging systems. In consumer photography and smartphone imaging, it may help improve image quality under low-light or high-noise conditions. In medical imaging, it could assist in recovering diagnostically useful details from noisy scans. In remote sensing, industrial inspection, and autonomous vision systems, it may improve the quality of images used for later analysis and decision-making. By combining denoising performance with detail preservation, the work offers a practical new solution for clearer and more reliable visual information extraction.