Shenzhen Link-Card Tech Co., LTDMenu

Интерфейс инфракрасного детектора и платформы обработки: FPGA, SoC и специализированный ISP

Технический блог · ISP для инфракрасных детекторов

Интерфейс инфракрасного детектора и платформы обработки: FPGA, SoC и специализированный ISP

Опубликовано 2026-09-05 · Link-Card Technology
Текст статьи доступен на английском языке.

Uncooled infrared detectors still emit raw video on vendor-specific timing, so the back end is an FPGA, an SoC with an infrared channel, or a dedicated infrared ISP. New packages move assembly out of the cleanroom. The published digital port still has to be checked part by part; it is not safe to assume MIPI CSI-2. This article takes the RTD6122C timing apart, then draws each back-end boundary from datasheets.

1. Classic detector interface timing: the RTD6122C as a sample

Access to visible-light CMOS image sensors (CIS) has long been standardized: a DVP parallel port or MIPI CSI-2 link plus I²C register writes — almost any SoC with a camera interface can drive one directly. Uncooled infrared detectors belong to an entirely different world: what they output is raw digital video on vendor-proprietary timing, with no unified electrical interface or protocol. Understanding this classic timing is the prerequisite for understanding how infrared back-end architectures evolved.

The sample used here is Raytron's RTD6122C: a new-generation Chinese detector with an effective array of 640×512 (full array 656×522), 12 μm pixel pitch, ceramic package, 8–14 μm spectral band, and 14-bit digital output — representative of the typical interface of today's Chinese mass-produced 12 μm parts.

Its interface works in five stages:

  1. Power-up and configuration: the back end first writes operating parameters to the detector over a serial link. The RTD6122 takes a configuration-header code on the SD0 line followed by 408 contiguous configuration bits (integration time, gain, frame rate, mirroring, defective-column replacement, etc.), effective on the next frame; a reset is 27 consecutive “1”s followed by one “0” on SD0. The physical form of the configuration link varies across vendors (single-wire serial encoding, SPI, etc.), but all are low-speed serial writes;
  2. Master clock input: the detector is driven by an external master clock whose frequency follows the line period and frame rate — the RTD6122 requires MCK ≥ 2770 × 522 × frame rate, about 72.3 MHz at 50 Hz; the ROIC (readout integrated circuit) performs per-pixel biasing, integration, and analog-to-digital conversion row by row;
  3. Synchronization and data output: sync signals are not necessarily dedicated pins — on the RTD6122, frame sync and line sync are specific in-band serial codes transmitted on SD0 (fixed patterns for FS and LS), with adjacent line-sync codes no closer than 2770 MCK cycles; pixel data is output 4-bit multiplexed on DO0–DO3, delivering 2 columns of 14-bit data every 7 MCK cycles (DO3&2 carry odd columns, DO1&0 even columns), with reference signals and the DTEMP/TEST columns embedded in the stream;
  4. On-chip correction with live write-back: the RTD6122 performs 6-bit on-chip non-uniformity correction (OOC); OOC data is continuously written back over SD0/SD1 (even columns on SD0, odd columns on SD1), takes effect in the same frame, and follows a 3-line pipeline — the detector–back-end link is bidirectional, real-time, and cannot be interrupted;
  5. The output is raw data, not images: two-point NUC, blind-pixel replacement, temperature compensation, digital detail enhancement (DDE), and pseudo-color mapping are all done by the back end. The picture coming straight out of the detector shows significant non-uniformity and noise and is not directly viewable.

A typical frame structure and signal relationships are shown in Figure 1.

RTD6122C interface timing (example): frame/line sync as in-band serial codes on SD0 MCK master clock (input) SD0 serial control Reset Config 408 bit FS code LS code LS code LS code (once per line) FS/LS are not dedicated pins but specific serial codes on SD0 (priority: reset > config > FS > LS) SD1 OOC write-back OOC 6 bit/pixel continuous write-back, effective in-frame (even columns on SD0) DO3–DO0 4-bit data (output) Flags + refs (repeat) 22 ref words Pixel data: 2 columns of 14 bit per 7 MCK DO3&2 odd columns / DO1&0 even columns After FS: 8 flag words + 14 reference signals repeat; after LS: 22 reference words once, then pixel data DTEMP (column 654) and TEST (column 655) embedded in the stream for temperature monitoring and debug MCK ≥ 2770 × 522 × frame rate (≈72.3 MHz at 50 Hz); config effective next frame, OOC in-frame Timing and coding per the RTD6122 uncooled IRFPA detector user manual V2.0 (Raytron, June 2025)
Figure 1. RTD6122C interface timing: frame/line sync as in-band SD0 serial codes, the 4-bit DO delivering 2 columns of 14-bit data every 7 master clocks, OOC continuously written back over SD0/SD1

The RTD6122C's key specifications are summarized in Table 1. Vendor differences in timing details deserve emphasis: whether sync is a dedicated pin or an in-band serial code, the data multiplexing ratio and bit width, the master-clock frequency formula, the physical form of the configuration link, even the definition of reference signals — all differ across vendors, and sometimes across batches of the same vendor. This is exactly why infrared detectors cannot be “plug and play” like CIS: for every new detector, the back end must re-adapt its drive and acquisition logic to the datasheet.

Table 1. RTD6122C key specifications (per the RTD6122 uncooled IRFPA detector user manual V2.0; defer to the datasheet for engineering use)
ItemValue (conditions)
Array sizeEffective 640×512 (full array 656×522, with DTEMP on column 654 and TEST on column 655)
Pixel pitch12 μm
Sensor technologyVOx microbolometer
Spectral band8–14 μm long-wave infrared
Digital output14-bit, line-by-line readout (4-bit multiplexed on DO0–DO3)
Frame rate≤60 Hz
Thermal sensitivityNETD <40 mK (CH/M variants, f/1.0, 50 Hz, 300 K; CR variant <50 mK)
Power consumption<170 mW (50 Hz, 25 °C, TEC excluded)
Radiometric range−40 to 600 °C (f/1.0)
On-chip correctionOn-chip 6-bit OOC (per-pixel output adjustment)
Operability≥99.5%
Operating temperature−40 to +85 °C
Package32-pin ceramic 22×22×4.58 mm (CR/CH), 40-pin metal (M)
InterfaceMCK master clock + SD0/SD1 serial control (in-band FS/LS codes) + 4-bit multiplexed DO0–DO3 output

2. The FPGA era: the default back end forced by proprietary timing

For a long time, no general-purpose SoC offered a standard interface that could take in the timing described above — the camera inputs of mainstream vision SoCs accepted only DVP and MIPI CSI-2, neither of which infrared detectors provide. FPGAs therefore became the default infrared back end: programmable I/O adapts to arbitrary proprietary timing, parallel pipelines carry infrared algorithms deterministically, and device choice and capacity leave ample headroom.

An FPGA serving as an infrared back end typically does five classes of work:

  1. Drive and configuration: generate the master clock and write parameters on that part’s configuration port. The port may be SPI, or an in-band code on SD0 as on the RTD6122. The power-up order is in the manual; a visible-light camera’s I²C sequence does not transfer.
  2. Data acquisition: deserialize and sample the multiplexed output in the source-synchronous clock domain (e.g., the RTD6122's 2 columns of 14-bit data every 7 MCK), then move it across clock domains into the system clock domain;
  3. Core correction: two-point NUC, blind/defective-pixel replacement, real-time temperature-dependent compensation, and in some scenarios shutter-based (FFC) calibration control;
  4. Image enhancement: AGC, DDE, noise reduction, pseudo-color mapping;
  5. Output adaptation: convert the result into standard interfaces such as BT.656/BT.1120, LVDS, USB, or Ethernet video.

This signal chain also explains where the development threshold of a “thermal camera core” lies: between the detector and viewable video stands an entire algorithm and logic chain that must be built in-house (Figure 2).

Infrared detector ROIC + focal-plane array raw data Back-end processing (FPGA / ISP / SoC) NUC · blind-pixel replacement · temp. compensation AGC · DDE · noise reduction · pseudo-color video Display / record / analysis BT.656 · LVDS · USB · Ethernet Application-layer control (parameters written on the serial port in the manual, not necessarily SPI)
Figure 2. The classic infrared imaging signal chain: the detector outputs raw data only — NUC, DDE, and the entire processing chain are borne by the back end

Early solutions were built almost exclusively on foreign FPGAs, with Xilinx (Spartan/Artix) and Altera (Cyclone) dominant. In recent years the model most frequently seen in thermal camera cores has become the Efinix Titanium-series Ti60: with logic capacity on the order of 60K, its cost, power, and package size cover exactly the demand scale of infrared back-end logic (typically tens of thousands of logic elements), displacing earlier devices in volume; Chinese FPGA vendors are entering this market in parallel. Beyond cost and power, supply-chain considerations and localized technical support drive this migration.

The FPGA route has equally clear costs: a high development barrier and a high bill of materials. Driver timing, clock-domain crossing, NUC/DDE algorithms, and DDR frame buffering all demand logic-development capability, and algorithm iteration cycles are measured in months; the device plus DDR, flash, and other peripherals costs significantly more than the later dedicated-chip solutions. This left room for the two routes that followed — SoC integration and the dedicated ISP.

Beyond full back ends, the market also has small-capacity FPGA/CPLD timing-adaptation solutions: low-cost, small programmable devices from vendors such as Gowin Semiconductor and Hercules Micro (Beijing) convert a detector's proprietary timing into standard DVP output, so that SoCs with camera interfaces can “read” the infrared data. These solve the interface-adaptation problem only; NUC, DDE, and the other algorithms remain the job of SoC software or a separate processing unit — a fit for products that reuse existing SoC platforms and do not demand a stringent image-quality chain.

3. SoCs with built-in infrared interfaces: the gap between claimed support and real-world performance

As the infrared market expanded, SoC vendors began building thermal-imaging input into their chips. The earliest and most successful is Shanghai Artosyn Electronics: its ARS31, released in 2022, integrated visible-light and thermal dual-spectrum engines and interfaces in one chip, with the public pitch of “eliminating the FPGA stage”; the later AR9341 and AR9481 series further claim direct drive of infrared detectors from 120 to 1920 pixels without an extra FPGA, targeting ultra-low power and ultra-low bandwidth. Among Chinese SoCs, this product line currently leads in infrared-support maturity.

Other vision-SoC vendors followed: HiSilicon's Hi3519DV500 states in its official specifications that it “supports access of mainstream CMOS-level thermal-imaging sensors”, with multispectral processing of thermal and visible light; Rockchip's infrared thermal imaging (IRFPA) solution announced in 2025 supports up to 16-bit data input with built-in OCC (per-pixel offset correction), automatic NUC, horizontal/vertical stripe removal, 3D/2D noise reduction, and dynamic range compression, and the new-generation vision chip RV1126B is among its infrared-supported devices; Weijing Technology's VS819 smart-vision SoC also claims infrared thermal-imaging access.

From an engineering standpoint, however, infrared support on SoCs other than Artosyn generally underdelivers, for three reasons:

4. Dedicated infrared ISPs: the chip class that genuinely lowers the barrier to using detectors

Between “full FPGA self-development” and “incidental SoC support” sits a more thoroughgoing route: the dedicated infrared ISP chip, which hardens the entire chain from detector drive timing through NUC and DDE to video output into a single chip, letting users integrate an infrared detector with a development experience close to “visible-light camera + ISP”. For vendors without an FPGA algorithm team, this directly and substantially lowers the barrier to building a qualified camera core.

The representative here is the T10 series of professional infrared ISPs from Jitai Microelectronics (Hangzhou) — a dedicated chip delivering the complete infrared image-processing chain, and one of the few professional infrared ISPs openly purchasable today (public information on this series is limited; the positioning described here comes from industry-channel information). A parallel data point comes from the detector vendors themselves: iRay Technology's in-house Falcon-series infrared ISP has iterated for several generations with industry-leading processing quality, but it is used only in iRay's own cores and cameras and is not sold separately. Raytron's automotive ASIC-ISP LY300, which passed AEC-Q100 automotive qualification in 2024, shows that dedicated infrared ISPs have reached automotive-grade reliability.

The dedicated ISP's limits mirror its strengths: a fixed chain means less flexibility than an FPGA, and deeply customized algorithms (specialized radiometry models, defense-specific requirements) hardly fit on it; but for civilian and industrial modules — the majority of the market — “low barrier + stable image quality” outranks “extreme flexibility”.

5. High-end systems: FPGA remains the irreplaceable back end

Dedicated ISPs and SoCs absorb the low-to-mid end and standardized demand; high-end infrared systems remain FPGA territory, with no sign of replacement in the short term:

The expected structure is layered coexistence: high-end custom builds stay on FPGA, standardized modules go to dedicated ISPs and SoCs, and the two routes evolve in parallel.

6. New packages and interfaces: assembly changed, timing may not have

The routes above all ask how the back end adapts to proprietary timing. A separate claim is that the detector itself should emit DVP or MIPI CSI-2, like a visible-light CIS. The direction is real. A specific part still has to be checked against its timing diagram; a packaging announcement is not an interface definition.

Raytron Microelectronics (a Raytron Technology subsidiary) published the OHLE3123 in March 2025. The public specification is 384×288, 12 μm, NETD <40 mK (f/1.0, 25 Hz, 300 K), power <65 mW at 25 Hz, thermal time constant <15 ms, 14-bit rolling readout, on-chip 9-bit non-uniformity correction, 36-pin SWLP, 10.4×10.4×4.3 mm³, weight <0.6 g. The input interface is listed as I²C (MODE0) or a 2-bit parallel port (MODE1). The public pages stress a double-layer window, assembly outside a cleanroom, and SMT. They do not include a MIPI CSI-2 or DVP timing diagram. SWLP lowers the cleanroom and placement barrier. It does not, on the published evidence, turn the output into a visible-light camera port.

AKIC Tech’s AK3812 is described publicly as a small package aimed at direct SoC drive. Until a timing diagram is public, that is an interface converging, not a demonstrated MIPI CSI-2 sensor.

Even a real DVP or MIPI output would only solve sampling. The raw data still needs NUC, DDE, and false color, on an SoC infrared channel of uneven depth or on a dedicated ISP. Package standardization and interface standardization are separate, and the public OHLE3123 specification currently settles only the first.

7. Five back-end architectures compared

Placing the routes above side by side, the differences in capability boundaries and suitable scenarios are summarized in Table 2.

Table 2. Infrared back-end architectures compared (qualitative summary; defer to each device's datasheet)
ArchitectureRepresentative devices/platformsRoleDevelopment barrierTypical applications
Full-custom FPGA back end Xilinx Spartan/Artix, Altera Cyclone, Efinix Titanium Ti60 Drive timing + acquisition + full NUC/DDE chain + output adaptation High (all logic and algorithms in-house) Large arrays, high frame rates, multi-sensor fusion, defense and deep custom builds
Small FPGA/CPLD timing adaptation Small FPGAs/CPLDs from Gowin Semiconductor, Hercules Micro Proprietary timing → standard DVP conversion Medium (little logic; algorithms on the SoC) Low-cost modules reusing existing SoC platforms
SoC built-in infrared interface Artosyn ARS31/AR9341/AR9481, HiSilicon Hi3519DV500, Rockchip RV1126B (IRFPA), Weijing VS819 Access + partial infrared processing (OCC/NUC/DDE, vendor-dependent) Low (software-driven) but bounded by support scope Bi-spectrum surveillance, UAV payloads, automotive assistance
Dedicated infrared ISP Jitai Micro T10 series (openly sold); iRay Falcon series (in-house only), Raytron LY300 (automotive) Drive + NUC/DDE + video output integrated in one chip Low (near visible-light camera development experience) Standardized core modules, automotive front-fit, industrial cameras
New package; interface still to be checked Raytron OHLE3123 (SWLP); AKIC Tech AK3812 OHLE3123’s published port is I²C or 2-bit parallel, not MIPI CSI-2; SWLP addresses assembly environment Timing still has to be checked; the algorithm chain stays on the back end SMT outside a cleanroom, small modules

Conclusion

In practice the split is: proprietary timing is still an FPGA job, and the Efinix Titanium Ti60 is now common in volume cores; Artosyn’s infrared channel is the more complete SoC implementation, while HiSilicon, Rockchip, and Weijing publish access claims whose image quality depends on how deep the vendor’s support goes for a specific detector; Jitai’s T10 is sold as a chip and iRay’s Falcon series stays inside iRay cores. Large arrays, high frame rates, and algorithms that must be modified stay on FPGA. The OHLE3123 moves assembly out of the cleanroom; its published interface is still I²C or 2-bit parallel. Selection means separating “can be soldered” from “plugs in as a MIPI camera”, then deciding whether NUC runs on an FPGA, an SoC, or a dedicated ISP.

Назад в технический блог

Нужна помощь с выбором или индивидуальным решением?

Link-Card помогает с выбором модулей, интеграцией и комплексными решениями.

Наш отдел продаж ответит на указанный email или телефон. Также можно написать напрямую на sales@link-card.com.