As a key technology for enhancing human visual perception at night, low-level light (LLL) night vision technology has been widely employed in military and civilian fields. Electron bombarded active pixel sensor (EBAPS) as a new generation of LLL night vision imaging devices has become a main research direction in the field of LLL night vision technology due to its low power consumption, high sensitivity, fast response, and excellent performance in extremely low illumination conditions (1×10-4lx). To detect and visualize optical signals in extremely low illumination conditions, LLL imaging devices should undergo a series of processes, including photoelectric conversion, signal amplification, and signal reading. However, the application of a series of amplification operations to the signal will inevitably result in noise amplification, which will significantly interfere with the original signal of the image. Therefore, it is vital to study the image denoising of LLL images. Compared to traditional image sensors such as complementary metal oxide semiconductor (CMOS), EBAPS is unique because it adopts bombarded electrons to achieve signal amplification. The signal amplification process in the method may result in a specific noise type, which is referred to as EBS noise in our study. The primary manifestation of EBS noise in the images is randomly distributed, oversaturated, and speckled pixel clusters. Although EBAPS shows great potential in LLL night vision field, the research on its complex mixed noises (EBS noise and Gaussian noise) suppression is still relatively rare. Traditional image denoising methods require highly complex mathematical reasoning and optimization processes while featuring poor performance under complex mixed noises. With the development of technology, deep learning-based image denoising algorithms have been widely adopted in optical imaging, including medical imaging, remote-sensing imaging, and mobile phone photography, due to their powerful feature extraction and excellent modeling capabilities.

Deep learning-based methods are divided into supervised and self-supervised image denoising algorithms based on whether a noisy-clean image-paired dataset is utilized during the training. In many supervised image denoising algorithms, based on limited clean datasets, different noise addition strategies have been employed to obtain noisy images such as additive white Gaussian noise, artificially synthesizing noisy-clean image pairs. However, artificial noise simulation of noise cannot accurately reflect the noise distribution in the objective world. In dealing with noise images in the real world, only relying on the model trained on artificial simulation noise may result in difficulties in yielding the desired denoising effect and accuracy. Therefore, considering that the noise in the EBAPS images is a mixture of EBS and Gaussian noises, the network architecture is designed and divided into two stages. In stage 1, the noise-noise paired dataset is constructed by adopting the iterative strategy and EBS noise, and thus the training phase does not rely on manually adding noise to construct the dataset. Additionally, the U-Net denoising model is built to realize the removal of EBS noise. In stage 2, based on the denoising results of EBS noise, a U-shaped blind-spot net drop model is designed and built for Gaussian noise for training to realize the removal of Gaussian noise.

The experimental data employed in our study are EBAPS images acquired in 1×10^-4 lx and 1100 V experimental conditions. We present a comparison of our algorithm with the benchmark classical denoising algorithms, including BM3D and the advanced self-supervised image denoising algorithms based on deep learning proposed in recent years. For BM3D, B2UB, and DBSN, the adaptability and denoising effect of these methods for EBS noise in EBAPS images are not ideal. However, although AP-BSN and MM-BSN methods can realize EBS noise denoising to a certain extent, they inevitably introduce significant side effects during denoising. This means that the image is over-smoothed, and the details and texture features of the image are lost, thus reducing image clarity and causing visual distortion (Figs. 9?12). In contrast, the proposed method successfully denoises EBAPS images and preserves the details and texture information of the original images. Compared with previous methods, the proposed algorithm yields the optimal performance of PSNR and SSIM, two key evaluation indexes (Tables 1 and 2). The experimental data and intuitive visual effects strongly demonstrate more sound performance and better results of the proposed method than those of current algorithms in targeting EBAPS image noise.

We propose a self-supervised two-stage convolutional neural network model for EBAPS images, which can maximize the preservation of image details and texture information while realizing the denoising of mixed noises (EBS noise and Gaussian noise). Additionally, the proposed method innovatively abandons the traditional practice of expanding the dataset with synthetic noise in the training phase, and instead directly utilizes the inherent noise characteristics of EBAPS images as the dataset. This strategy not only reduces the reliance on synthetic noise but also motivates the proposed denoising algorithm to capture and generalize the complexity of EBAPS image noise more effectively. Our experimental results show that the proposed method achieves better performance than state-of-the-art algorithms on the industry-recognized image quality evaluation metrics PSNR and SSIM. However, there is room for improvement in further optimizing the preservation of image details and simplifying the network structure, which is the main direction of our future research.