基于像素暗噪声估计的 EBAPS 图像自适应小波阈值降噪

Objective Electron bombarded active pixel sensor (EBAPS) is a kind of high-performance low^-light video imaging device with vacuum^-solid mixture. Since the domestic EBAPS is still in the early stage of research, various noises are unavoidable during the imaging process. However, the classical denoising algorithms, such as total variation, wavelet, and various edge-preserving filters, are aimed at additive white Gaussian noise (AWGN) with constant standard deviation, and the noise variance level should be known. Since the noise of EBAPS is a mixture of dark noise, shot noise, and fixed pattern noise, with unknown noise level, the denoising algorithms designed for AWGN are not effective for EBAPS images. Therefore, we first analyze the noise characteristics of EBAPS, including AWGN independent of signal strength, Poisson noise varying with signal strength, and fixed noise. Then, we propose a noise estimation method for a single frame image by employing dark pixel structure characteristics of EBAPS. Finally, the traditional wavelet threshold denoising is improved according to the estimated noise intensity, and an adaptive variable threshold is put forward according to the noise intensity of the single frame image. We hope that our denoising method can improve the low^-light image quality of EBAPS with lower computation and less frames. Methods The proposed algorithm includes noise estimation and wavelet denoising. The noise estimation includes the following steps. First, the properties of the solid^-state imaging device and vacuum imaging device are combined to infer the noise source of EBAPS, and the relationship between EBAPS noise and signal intensity is obtained by experiments using the photo transfer curve (PTC) method. Then, based on the dark pixel structure of EBAPS, we infer the unified noise intensity model using a single frame image. The wavelet denoising decomposes the image into multiple sub^-bands at different resolutions and scales, the image subject information exits in the low^-frequency sub^-band, and the noise and detail information exits in the high-frequency sub^-band. By setting a threshold on the coefficients of high-frequency sub^-bands, the noise can be almost removed. Finally, the denoised image is restored by inverse wavelet transform. The performance of wavelet threshold denoising depends on the threshold. A larger threshold will shrink the signal features to result in image over-smoothing and create blur and artifacts, while a smaller threshold will leave more noise information. Since the images of EBAPS have low signal-to^-noise ratio and complex noise sources, the classical thresholds such as UT threshold, Rigrsure threshold, and Min^-max threshold are not effective. Thus, based on wavelet threshold denoising, according to image noise intensity estimation in the previous step, adaptive wavelet threshold denoising based on pixel dark noise of EBAPS (AWT-PDN) is designed. Results and Discussions The unified noise intensity model [Eq. (2)] can be divided into three stages. In the first stage, when the signal intensity of the light pixel region is less than that of the dark pixel region, the noise is mainly AWGN and basically remains constant. In the second stage, the signal intensity of the light pixel region increases and the noise intensity rises gradually, which follows the Poisson distribution. In the third stage, when the signal intensity of the light pixel region increases to a certain value, according to the histogram distribution shape and Poisson distribution theorem, the noise at this time is close to the Gaussian distribution, with the intensity remaining constant. On the other hand, the image processing results under different illuminance of wavelet threshold denoising with different thresholds are shown in Figs. 11 and 12. Subjectively, the noise of images under all illuminance is suppressed by selected methods and the image noise of our AWT-PDN method is less in the bright area of the image. For a more intuitive observation of the detail of images, a line crossing the black^-white edge in the image is emphasized in pixel value (Fig. 11). In 1×10^-2 lx and 5×10^-3 lx illuminance conditions, the line edge processed by our AWT-PDN method is clearer and smoother, but that processed by other selected thresholds has more burrs. Objectively, PSNR [Eq. (18)], SSIM [Eq. (19)], and AFD [Eq. (20)] are employed to evaluate the proposed method. As shown in Table 2, the proposed AWT-PDN method has better performance than others and can preserve edges. Conclusions The noise sources of EBAPS are analyzed, including Gaussian noise independent of signal strength, Poisson noise varying with signal strength, and fixed pattern noise, and the relationship between EBAPS noise and signal intensity is obtained by experiments using the PTC method. Then, based on the dark pixel structure of EBAPS, we propose an adaptive wavelet threshold denoising method AWT-PDN for EBAPS images based on pixel dark noise. According to the noise intensity distribution of the EBAPS single frame image, an adaptive variable threshold wavelet threshold denoising method is obtained. Experiments show that the proposed AWT-PDN method can reduce the EBAPS imaging noise, and yield a better noise reduction effect than traditional threshold methods in 5×10^-3 lx illuminance conditions.