Advances in imaging and optical technologies have greatly enriched the physical information captured and persevered by multiple sources and dimensions of images. Polarimetric imaging, leveraging measurements of polarimetric parameters that encode distinct physical properties, finds wide applications across diverse domains. However, some critical polarization information is highly sensitive to noise, and denoising polarimetric images while preserving polarization information remains a challenge.

Recently, Prof. Haofeng Hu, along with Assistant Researcher Hedong Liu from Tianjin University, focused on the development of polarimetric image denoising technology, published a review paper titled "Review of polarimetric image denoising" in Advanced Imaging. This review offers a comprehensive overview for diverse methods that have advanced the domain of polarimetric image denoising. These methods are first classified as learning-based and traditional methods. Then, the motivations and principles of different types of denoising methods are analyzed. Finally, some potential challenges and directions for future research are pointed out.

This review analyzes the sources of noise in polarization imaging and summarizes the mathematical principles and transmission amplification characteristics of noise in the imaging process based on the physical process of photoelectric conversion. Through simulation analysis, the high noise sensitivity of key parameters such as the Degree of Polarization (DoP) and Angle of Polarization (AoP) is verified. The analysis indicates that even low levels of noise can lead to significant deviations in the measured values of polarization parameters. Therefore, improving the quality of polarization images to ensure the accuracy of DoP and AoP is crucial for the further development and application of polarization imaging technology.

The diagram of the noise formation model.

In recent years, deep learning-based polarization image denoising methods have rapidly developed and demonstrated great potential in handling noise-sensitive polarization parameters. This paper thoroughly reviews the strong synergy between polarization imaging and deep learning technologies. The integration of optical information can impart clearer physical meaning to deep learning techniques and inject new vitality into their advancement. Therefore, the paper analyzes the role of optical technologies, such as infrared imaging and 3D imaging, in promoting deep learning-based polarization image denoising methods. Conversely, deep learning improves the overall performance of polarization imaging technology through a data-driven approach. By employing various deep learning techniques, such as supervised learning, self-supervised learning, and transfer learning, obtaining high-quality polarization images and information becomes simpler and more accessible, significantly advancing the application of polarization imaging in numerous fields. The organic integration of deep learning and polarization imaging technologies could offer solutions to various challenges and difficulties in both domains.

Synergy between polarimetric imaging and deep learning techniques.

Noise degrades the quality of polarization images and hinders the further application of polarization imaging technology, making denoising techniques a key issue in this field of research. This paper systematically reviews polarization image denoising techniques, introducing both traditional denoising methods and deep learning-based approaches, while analyzing their main characteristics and key features. It also highlights potential future research directions and current challenges. Although deep learning techniques, particularly convolutional neural networks, have shown great potential in polarization image denoising tasks, practical applications still need to consider the integration of physical and polarization information to improve denoising performance and interpretability in handling complex real-world noise. Furthermore, obtaining paired polarization datasets is challenging, so exploring methods that reduce data dependency, such as unsupervised and self-supervised learning, is one of the future directions. The use of multi-sensor fusion methods, such as combining infrared and color images, can further enhance the quality of polarization images, providing an effective solution for obtaining high-quality images.