Noise is one of the main causes of image degradation in photon-starved environments, particularly in applications requiring short exposure time, such as ultrafast imaging and low-light imaging. By manifesting in the information capturing process, the noise is able to degrade significantly the subsequent image analysis process and impede the intended applications. Consequently, severe noise could flood the information, thereby deteriorating targeted detection and recognition. Therefore, developing effective strategies to minimize or at least suppress the noise is highly desirable.
Among the existing optical imaging technologies, polarimetric imaging has drawn significant research attention. It is based on measuring the spatially resolved polarization state of light, which is widely used in different fields, including target detection, three-dimensional (3D) reconstruction and image dehazing. In most practical applications, essential polarimetric information considered is the angle of polarization and degree of polarization. However, polarimetric information involves nonlinear operations and is highly sensitive to noise, which could interfere with its applications. While removing noise to accurately restore polarization information is of extreme importance in such applications, it has remained a challenging task that requires more effective solutions.
Through convolution, it is possible to extract features by blending spatial and cross-channel information. As a result, the convolutional neural network (CNN) has played a key role in improving the performance of denoising, which has proved to be more effective than non-data-driven methods like block-matching and 3-D filtering algorithm. Interestingly, CNN has also been identified as a powerful tool for polarimetric image denoising. However, it lacks the flexibility needed to deal with complex noisy images since all channel-wise features are treated equally. As a result, it is important for denoising models to selectively emphasize on informative features while suppressing corrupted and less useful ones.
Herein, a team of Tianjin University researchers: Dr. Hedong Liu, Professor. Yizhu Zhang, Professor Zhenzhou Cheng, Professor Jingsheng Zhai and Professor Haofeng Hu developed an attention-based residual neural network for improved polarimetric image denoising. It was specially designed to overcome the challenges associated with existing image denoising methods and allow effective noise removal and restoration of polarization information of polarimetric images. To accomplish this, a new channel attention mechanism was adopted. In addition, an adaptive polarization loss was designed to help in guiding the network in focusing on desired polarization information. Their work is currently published in the research journal, Optics Letters.
The research team demonstrated the effectiveness of the newly introduced channel attention residual dense block in extracting the important and useful informative features while suppressing less useful ones. To this end, this method was effective in suppressing the noise polarization information in the images, even in practical situations involving severe noise. The experimental results were compared to those obtained from existing methods. As anticipated, it produced better performance and outstanding results. Moreover, it clearly revealed the underlying channel attention mechanism that contributed to its success. In particular, the underlying channel mechanism made the learning process easier and more explicit by visualization of informative features.
In summary, the authors demonstrated successfully the effectiveness and applicability of attention-based neural network in removing noise and restoring polarization information of polarimetric images. The experimental results confirmed the validity of this method and its superiority over existing methods. In a statement to Advances in Engineering, Professor Haofeng Hu stated that the new attention-based neural network will contribute to broadening the polarimetric imaging applications, especially in scenarios involving low signal-to-noise ratios like ultrafast and low light imaging.
Liu, H., Zhang, Y., Cheng, Z., Zhai, J., & Hu, H. (2022). Attention-based neural network for polarimetric image denoising. Optics Letters, 47(11), 2726–2729.