With the continuous improvement in complementary metal-oxide-semiconductor (CMOS) manufacturing technology, the performance of CMOS sensors has been significantly enhanced, making them comparable to charge-coupled device (CCD) sensors. Additionally, due to their advantages such as high integration, small size, low power consumption, and fast speed, CMOS sensors have gradually become the primary imaging devices in many fields such as aerospace, biomedical, industrial vision, and digital photography. In optical remote sensing imaging, the development of high-performance optical CMOS cameras requires significant investment and long development cycles. Adopting computer technology to simulate remote sensing images is of great significance for camera design, image quality assessment, and research on data processing algorithms. Building noise simulation models that match the working principles of optical sensors is a necessary and meaningful task. Currently, there are many studies related to noise simulation of CCD and CMOS sensors, but research on noise simulation of area array CMOS sensors often focuses on one or a few types of noise. Therefore, it is necessary to comprehensively analyze the noise data characteristics of area array CMOS sensors and build a complete noise simulation model.

We analyze various sources of noise during the imaging process based on the physics of CMOS sensors, build a set of noise models for CMOS sensor images and conduct noise simulation and validation of the effects. Firstly, the noise generated during the signal conversion of CMOS sensors is analyzed. Then, based on the noise characteristics, we build a noise model and propose a parameter calibration method for this model. Finally, we carry out simulations based on the calibrated noise parameters and evaluate and compare the simulated images under different noise models with actual images. By employing aerial photography data from Huizhou, Guangdong Province captured by a CMOS camera array, we adjust the noise model parameters to simulate the main noise components, thus validating the reliability of the noise simulation model and the effectiveness of the parameter calibration method.

To verify the superiority of the proposed noise model, we conduct laboratory simulations using two different noise models. Lab-captured R-channel bright field images with grayscale values saturated at 10%, 25%, 45%, and 60% of the dynamic range are selected as experimental objects. The simulated images using the proposed noise model exhibit brightness and noise distributions similar to those of the real images [Figs. 9(a) and 9(b)]. The average relative deviations of SNR, PSNR, and SSIM between simulated images using the Poisson-Gaussian distribution mixed noise model and real images are 4.83%, 1.21%, and 1.59% (Table 2). Meanwhile, the average relative deviations of SNR, PSNR, and SSIM between simulated images using the proposed noise model and real images are 3.85%, 0.99%, and 1.33% (Table 2). It is observed that the proposed noise model has smaller relative deviations, which validates the effectiveness of the proposed method. Furthermore, the result comparison under different grayscale values using the same noise model indicates that under identical camera settings, higher light intensity leads to higher SNR. Additionally, SNR increases more rapidly at lower illuminations and transitions to a relatively slower growth rate at higher illuminations, which conforms to the general rule of SNR varying with signal intensity. The camera is primarily affected by photon shot noise at low illuminations, while at higher illuminations, it is mainly influenced by residual noise such as read noise. This may explain the relatively larger relative deviation when the grayscale value is saturated at 45%. To verify the influence of each noise component in the noise model on the image, we make adjustments to the parameters of the noise model based on aerial photography data captured by a CMOS camera. As the total gain increases, the image photon shot noise also rises to decrease image contrast, which is particularly noticeable in darker areas [Figs. 11(a) and 11(b)]. Row noise introduces noticeable horizontal stripe distortion in the image, and as the scale parameter of row noise increases, the horizontal stripes become more pronounced [Figs. 12(a) and 12(b)]. Increasing the scale parameter of read noise leads to a decrease in image sharpness, especially in low-contrast areas. Noise blurs edges and details in the image [Figs. 13(a) and 13(b)].

We propose a comprehensive noise simulation model, which fits noise data with heavy-tailed distributions using the Tukey lambda distribution. Meanwhile, we provide a method to calibrate model parameters for the noise model. By applying this noise model and calibration method to a scientific-grade CMOS camera, we obtain the noise model parameters for the camera sensor. Simulations are then conducted based on the calibrated noise parameters, with the simulated images compared with real images. The results show that the average relative deviations of SNR, PSNR, and SSIM between simulated and real images are 3.85%, 0.99%, and 1.33% (Table 2). Additionally, the simulated images using our proposed noise model exhibit smaller relative deviations than those using different noise models. By adopting aerial photography data captured by a CMOS camera, we adjust the noise model parameters to simulate the main noise components. The simulation results reflect the influence of changes in noise model parameters on images and are consistent with the theoretical trends of noise characteristics. These two simulation experiments validate the reliability of the noise simulation model and parameter calibration method, and the effectiveness of the simulation method.