Low-level-light night vision technology enables the generation of visible-light images that are clearly recognizable to the human eye under low-illumination conditions. An electron bombarded active pixel sensor (EBAPS) represents a vacuum-solid hybrid low-level-light imaging device that delivers exceptional performance in sensitivity, resolution, power efficiency, and compact design, making it valuable for low-light-level night vision applications. However, the imaging procedure introduces complex noise patterns. The electron bombardment process used for signal enhancement generates electron bombarded semiconductor (EBS) noise, characterized by spatially random distribution, multi-pixel aggregated clusters, and a diffuse pattern that gradually darkens from center to edges, manifesting as bright spots. While traditional denoising methods effectively address common Gaussian noise and salt-and-pepper noise that occur in solid-state imaging devices, they prove inadequate for EBS noise removal. Currently, effective methods for EBS noise removal remain limited. Previous attempts combining noise detection with median filtering yielded unsatisfactory results. Processing speed represents another critical factor in practical applications, with some complex algorithms requiring extensive processing time. Deep learning-based approaches show promise in handling unconventional EBS noise but face challenges in dataset construction, processing speed, and generalization capability. Therefore, we propose a Harris-guided adaptive switching median and bilateral filtering (HASMBF) algorithm for rapid and effective EBAPS image noise removal.

The proposed algorithm comprises three main components: EBS noise detection, adaptive switching median filtering, and bilateral filtering. The Harris corner detector demonstrates effectiveness in EBS noise detection. However, it exhibits over-detection tendencies, where non-noise pixels surrounding EBS noise pixels are incorrectly classified as noise pixels. These misclassified pixels typically display significant brightness differences from actual noise pixels. To address this limitation, we incorporate the Otsu threshold to optimize the Harris corner detector results, developing a Harris-Otsu joint noise detection algorithm. The Otsu method functions as an adaptive thresholding technique based on image histogram statistics. Its fundamental principle maximizes between-class variance, enabling automatic computation of optimal segmentation thresholds based on image grayscale distribution, thus effectively separating targets from backgrounds without manual intervention. The process begins with Harris corner detector identifying approximate EBS noise regions, followed by retention of only pixels exceeding the Otsu threshold as final detection results. Subsequently, an adaptive switching median filter guided by detection results removes EBS noise. Finally, a bilateral filter eliminates residual Gaussian and Poisson noise.

The proposed Harris-Otsu joint noise detection algorithm accurately detects EBS noise pixels (Fig. 4). To verify the denoising performance of the HASMBF algorithm, denoising experiments were subsequently conducted using both simulation images with artificially added noise and real EBAPS images. The simulation images are generated by adding artificial EBS noise and Gaussian noise, and are divided into three categories based on illumination levels (Fig. 6). The real EBAPS images were captured using an EBAPS camera under illumination condition of 5×10^-4 lx (Fig. 10). Experiment results show that the algorithm proposed in this paper is effective in removing the mixed noise in EBAPS images (Figs. 7, 8, 9 and 11). EBS noise is significantly removed after being processed by the proposed algorithm. Our method also performs well in both PSNR and SSIM (Tables 1, 2, 3 and 5), and it also attains a competitive processing speed (Tables 4 and 5). For real EBAPS images, the proposed algorithm improves PSNR by approximately 10 dB and achieves an SSIM above 0.9. Meanwhile, the processing time is only 1.8 times that of the classic bilateral filtering algorithm.

This study analyzes the imaging principle and noise characteristics of EBAPS, including Gaussian noise and EBS noise. The proposed HASMBF algorithm demonstrates accurate EBS noise detection and effective mixed noise removal in EBAPS images. The denoising performance evaluation utilizes both simulation and real images. Experimental results confirm the method’s effectiveness in noise suppression while maintaining efficient processing speed. However, the algorithm requires parameter adjustments based on environmental variations, including noise detection threshold and bilateral filtering kernel parameters. Further research should focus on developing an adaptive parameter-setting mechanism to enhance the algorithm’s environmental adaptability.