The optical design was tailored to match a 41-cm aperture f/2.5 telescope of Sun Yat-sen University, as shown in Figure 2. Given the relatively higher readout noise and dark current of the detector compared to leading international detectors, minimizing additional noise sources—particularly stray light and optical efficiencies—is crucial to fully exploit the exceptionally low infrared sky background at Dome A. To maximize optical throughput in H band while suppressing off-axis thermal radiation and unwanted reflections, a cold shield and a custom-designed H-band filter were incorporated, significantly limiting background flux and enhancing image contrast.

The camera’s FOV reaches approximately 0.5 × 0.4 square degrees (f/0.6), fully utilizing the telescope’s image field without significant vignetting or peripheral aberrations. Figure 3 illustrates the transmittance curves of the camera’s filter and window at room temperature (300 K), compared with atmospheric transmission.

The camera employs a 640 × 512 MCT focal plane array developed by SITP (see Figure 4a). The detector features a pixel pitch of 15 µm and a cutoff wavelength of 2.0 µm, covering the full H-band atmospheric window. The detector adopts a hybrid structure in which a MCT layer is bump-bonded to a silicon readout integrated circuit (ROIC) chip with indium pillars (see Figure 4b). This hybrid integration method provides robust mechanical performance and reliable electrical interconnectivity under thermal cycling conditions (Zhang et al., 2012). Additionally, careful thermal management during cryogenic characterization tests has demonstrated stable detector performance at low temperatures (Jiménez et al., 2020; Stoyanov et al., 2019).

In the H-band camera, the optics and detector components were integrated into a vacuum-sealed Dewar, which is cooled to approximately 80 K by a Stirling-cycle cryocooler (see Figure 4c). The Dewar employs an aluminum nitride (AlN) mounting plate for both mechanical support and efficient thermal conduction. Electrical connections to the external readout electronics are made via gold wire bonds and routed through hermetically sealed feedthroughs.

Given the extremely low ambient temperatures at Dome A, typically below −60°C, a heating film is integrated into the system to maintain the refrigeration unit’s operational range above −40 °C. To prevent thermal infrared radiation from the heater affecting the detector’s photosensitive surface, a radiation isolation ring is installed between the Dewar window and outer shell, effectively blocking internal thermal emissions from reaching the focal plane.

The readout circuit in the MCT detector is a capacitive transimpedance amplifier (CTIA) architecture, which is operated in reduced bandwidth and incorporated two selectable integration capacitance modes for gain selection, 5 fF and 10 fF (see Figure 5). With reset after integration time, the response electrons of each pixel are integrated and converted into voltage signals. After sampling, the signals are selected by rows and columns, and the array pixel signals are transmitted to the output stage one by one, and finally output outside the chip.

The electronics of the camera consists of a Main Control Board (MCB) and a Power Board (PB). The MCB handles system control, readout, and data processing, while the PB provides power and manages temperature control through a heating module. The functional diagram of the H-band camera is shown in Figure 6. To enhance operational robustness under remote and harsh conditions, the camera supports dual communication interfaces (UDP and USB 2.0). The UDP interface is typically preferred due to its longer transmission distance, but the USB link provides a redundant fallback channel in case of network disruptions. The total power consumption including the cryocooler and all electronics, is approximately 45 W. This low-power design is particularly well-suited for autonomous and long-term deployment at Dome A, where both power availability and thermal balance are critical constraints.

The detector signal passes through an ADC driver with a gain of 1.667—selected to match the input dynamic range—and is sampled by an 18-bit ADC. An FPGA then transmits the upper 16-bit data to the host computer (see Figure 7) for maximizing dynamic range.

Comprehensive performance characterization is essential for scientific-grade infrared cameras, especially when deployed in extreme environments such as Antarctica (Sutherland et al., 2015; Zhang et al., 2019). These tests provide quantitative validation of the camera’s readiness for high-fidelity imaging under low-background infrared conditions (Tubiana et al., 2015).

A dedicated test platform was established for this purpose, as illustrated in Figure 8. A critical component of the test setup is the blackbody radiation source, which directly influences the accuracy of detector calibration (Qu et al., 2025; Wright et al., 2015). The blackbody source features a 100 mm emission surface, temperature control precision of ±0.02°C, resolution of 0.01°C, and temperature stability better than 0.1°C per 10 min. During testing, the camera was precisely aligned so that all pixels viewed only the blackbody surface, thereby minimizing stray radiation. Under conditions of assumed uniform blackbody emission, the illumination non-uniformity across the detector’s field of view was measured to be ∼3.74%.

The photon transfer curve (PTC) method was employed to evaluate gain for both the central 40 × 40 pixels region and the entire detector area (see Figure 9) (Sun et al., 2022). For the 5 fF integration capacitor, the measured gains are 1.04 e⁻/ADU and 1.13 e⁻/ADU in these two different area, where ADU denotes an analog-to-digital unit. At a setting of 5 fF, the measured readout noise is approximately 29.8 e⁻, and the readout electronics contributed about 5.5 e⁻, indicating that the main noise component is inside the detector. According to the standard of response rate less than 50% and greater than 200%, the proportion of bad pixels is 0.44%

Inter-pixel capacitance (IPC) estimation is based on the spatial autocorrelation function (ACF) approach (Moore et al., 2004). We first subtract the mean from each image to obtain zero-mean data, and then calculate the two-dimensional ACF using Fast Fourier Transform (FFT). The central peak of the ACF is normalized to unity, and the IPC is estimated by averaging the autocorrelation values of the four nearest-neighbor pixels adjacent to the central peak, as shown in Figure 10. The derived IPC value was below 0.01, suggesting a negligible influence on gain estimation. Therefore, IPC correction was not applied in this analysis. While IPC may affect the Point spread function (PSF) in precision photometry, accurate correction would require a full-system model including optical aberrations and on-sky PSF measurements (Kannawadi et al., 2016), which are outside the scope of the present laboratory characterization and are planned for future work.

Since the detector was already sealed within a Dewar, conventional dark current measurement using a cold shutter is not feasible. As an alternative, the camera was placed in a refrigerated environment to reduce the ambient photon flux, enabling an approximate evaluation under dark-field conditions. At a stabilized temperature of −25°C, the longest exposure time of the camera exceeds 20 min, and the incident background in the H-band was estimated to be approximately 9 photons per second per pixel, while the measured signal slope was approximately 17 electrons per second per pixel. This value reflects the combined contributions from the detector’s intrinsic dark current, residual ambient background radiation, and readout noise.

PRNU was measured to be 4.44% using flat-field illumination at various exposure times. The Signal-to-Noise Ratio increases markedly with higher illumination, reaching a maximum of about 135 (42.60 dB) before the detector approaches full well capacity. Based on a comparative analysis with the 2MASS system, which achieved a limiting magnitude of H ≈ 15.1 at SNR = 10 using a 1.3 m f/3.5 telescope and a ∼7.8 s exposure per scan, we estimate the limiting magnitude of our 0.41 m f/2.5 system under Dome A conditions to be approximately 12.8 mag for a 10 s exposure at the same SNR level. With frame stacking or extended exposures up to 60 s, the limiting magnitude would be expected to improve to 14.0–14.3, demonstrating the system’s capability to detect faint astronomical sources under background-limited conditions.

In an ideal PTC measurement, it is assumed that each pixel in the detector follows the same Poisson distribution for every acquisition. In reality, the mean photon count for each pixel is not identical. For pixel i, its Poisson mean is λ i , which can vary depending on the pixel’s FOV angle—pixels with a larger FOV angle collect more photons, and vice versa.

Consequently, if one treats all pixels as a single ensemble when estimating the “variance vs mean,” this is equivalent to randomly selecting a pixel i with probability 1/N (where N = 640 × 512), and then drawing a single observation from Poisson ( λ i ). From the perspective of mixed distributions, the random variable X m i x (formed by mixing measurements across all pixels) has the following expectation and variance, refer to the appendix for details. Equations 1, 2 show the mathematical expectation and variance of X_mix , respectively: E X m i x = 1 N ∑ i = 1 N λ i , (1) V a r X m i x = 1 N ∑ i = 1 N V a r P o i s s o n λ i ⏟ M e a n P o i s s o n v a r i a n c e + V a r λ i ⏟ N o n u n i f o r m d i s t r i b u t i o n    = 1 N ∑ i = 1 N λ i + V a r λ i , (2)

For a Poisson distribution with parameter λ i , the variance is λ i Hence the first term represents the average of the Poisson variances of the pixels, while the second term is the “extra mixture variance” caused by differences in the pixel means. Defining Equation 3 as follows, λ ¯ = 1 N   ∑ i = 1 N λ i , σ λ 2 = V a r λ i , (3) we obtain the simplified variance formula in Equation 4, v a r X m i x = λ ¯ + σ λ 2 , (4)

Accordingly, we define the variance-to-mean ratio in Equation 5 as: v a r X m i x X m i x ¯ = λ ¯ + σ λ 2   λ ¯   = 1 + σ λ 2 λ ¯ , (5)

The standard PTC method assumes all pixels follow Poisson ( λ i ) with the same λ . Under this assumption, v a r X m i x ≈ g × X m i x , and a linear fit to v a r X m i x  vs .   X m i x yields the slope g . In practice, though, the slope is (Equation 6) becames: S l o p e = v a r X m i x X m i x ¯ = g + σ λ 2 λ ¯ g   >   g unless all  λ i   are identical , (6) leading to a fitted gain higher than the detector’s true gain.

The photon transfer curve method is employed to calculate pixel gain by analyzing the linear relationship between the mean signal and variance of flat-field images at different exposure times. This method involves capturing 400 × 50 flat-field images, where 400 denotes the number of frames taken for each exposure time t i and 50 represents the number of different exposure times.

For each pixel x , y , the mean signal S x , y t i and noise variance σ x , y 2 t i are calculated using Equations 7, 8, respectively: S x , y t i = 1 n ∑ j = 1 n I x , y t i , j , (7) σ x , y 2 t i = 1 n − 1 ∑ j = 1 n I x , y t i , j − S x , y t i 2 , (8)

Where I x , y t i , j is the signal value of pixel x , y in the jth frame under exposure time t i .By plotting σ x , y 2 t i against   S x , y t i , a linear PTC response is obtained.

Pixel-level gain testing was conducted across two experimental setups, and obtained a flat and regular gain distribution on the testing platform with an optical system, which conforms to the gain inconsistency phenomenon caused by CMOS readout circuit process deviation and has obvious horizontal and vertical patterns. For the test results of H-band cameras without optical lenses, the horizontal and vertical distribution of gain cannot be observed, highlighting the central circular distribution, as shown in Figure 11. The pixel level gain has been normalized, according to Equation 9 and the gain calculated based on the PTC is set to 1. g a i n N o r m a l i z e d = g a i n P i x e l − l e v e l / g a i n P T C , (9)

The heatmap in Figure 11 reveals a consistent pattern: the central region of the detector exhibits lower normalized gain values, while higher values are observed toward the periphery. Importantly, since the PLG method relies on photon shot noise statistics and is independent of the absolute illumination level (assuming Poisson-distributed incident light), these spatial variations cannot be attributed to vignetting, stray light, or illumination gradients. To further investigate this trend, as shown in Figure 12a, the x-axis represents coordinates from the center, and the y-axis represents the g a i n N o r m a l i z e d . There are 95% of pixel-gain points between the red and green lines. The gain distribution at a distance of 200 pixels from the center as shown in Figure 12b. Kernel Density Estimation (KDE) is a non-parametric method used to estimate an unknown probability density function (Silverman, 2018), the distribution is basically consistent with a Gaussian distribution.

Testing systems with optical lenses have more regular pixel level gains for detectors, while detectors in H-band cameras may not receive uniform incident light due to different fields of view, which may affect the measurement results of pixel level gains. We propose a geometric FOV compensation method to mitigate pixel-level gain non-uniformity.

Pixel A is a point on the detector, as shown in Figure 13a, and the conical surface formed by point A along the cold screen hole is the solid angle at which pixel A can receive radiation, Figure 13b shows the cross-section along the XOZ plane. Based on the cross-section, AE is the angle bisector of angle BAC, and point E is the intersection point of this angle bisector and the cold screen hole plane. Therefore, based on geometric relationships, we calculate parameters e, f, and s using Equations 10–12, respectively e = b c b + c 2 − a 2 / a + c , (10) f = 2 b 2 + c 2 + a 2 / 2 , (11) s = g − e 2 − h 2 , (12)

We convert the oblique elliptical cone surface of the solid angle of point A into a regular elliptical cone surface with AE , as shown in Figure 13c and the lengths of its major and minor axes can be calculated according to Equations 13, 14 a e l l i p s e = a / 2 2 − s 2 , (13) b e l l i p s e = e   tan A / 2 , (14)

Calculate the solid angle numerically based on the height of the regular elliptical cone and the length of the major and minor axes of the ellipse (Li et al., 2012), as described by Equation 15. Ω = ∫ d S r 2 , (15)

The solid angles of light windows vary across different pixel locations on the detector, resulting in a discrepancy in the number of photons received by pixels at different positions. We consider the central area of the detector as the most important observation area, therefore the circular area at the center is designated as area A, the circular area to the X-axis boundary is designated as area B, and the surrounding corners are designated as area C. As shown in Figure 14, it can be observed that the variation in the solid angle is minimal within the central region (Region A), while it decreases significantly as the pixels approach the edge of the detector. The normalization and subsequent pixel-level gain corrections are implemented using Equations 16, 17: The normalized field-of-view (FOV) coefficient and the parameter for pixel-level gain correction are calculated using Equations 16, 17 respectively: F O V N o r m a l i z e d = F O V   V a l u e / F O V   V a l u e ¯ , (16) g a i n N o r m a l i z e d   C o r r e c t e d = g a i n N o r m a l i z e d × F O V N o r m a l i z e d , (17)

The pixel gain was corrected according to Equation 18, yielding a curve representing the average gain as a function of the distance from the center of the detector, as shown in Figure 15a. In regions A and B, the corrected gain remains approximately constant at a value of 1 (∼ g a i n P T C ). After using FOV correction, the detector gain is basically consistent in the central area, close to the overall calculated gain value, while the gain value in the corner area is smaller. C o r r e c t e d g a i n = g a i n N o r m a l i z e d   C o r r e c t e d × g a i n P T C , (18)

According to the calculation of the average values of pixels before and after correction based on three partitions, as shown in Figure 16, it can be concluded that using pixel level correction can make the spatial distribution of gain more uniform. This indicates that the different fields of view of the detector may have a correlation with the pixel level gain calculation method.