A downhole extremely low-illumination dual-light fusion passive physical enhancement imaging method and device

Abstract

The application relates to the technical field of mine safety monitoring and optical imaging, and discloses a downhole extremely low-illumination dual-light fusion passive physical enhancement imaging method and device, which comprises a passive dual-light collection module, a real-time preprocessing module, a hierarchical fusion processing module and an intelligent identification output module. The method synchronously acquires micro-light and infrared images under the condition of no active light compensation, improves the signal-to-noise ratio through physical enhancement and non-uniformity correction, fuses pixel-level, feature-level and decision-level information, and realizes low-power-consumption and high-reliability target identification and output in combination with a light-weight AI model. The application can effectively solve the problems of poor imaging quality, difficult target identification and high system power consumption in the extremely low-illumination environment downhole.

Description

Technical Field

[0001] This invention relates to the field of mine safety monitoring and optical imaging technology, specifically to a method and device for passive physical enhancement imaging with dual-light fusion in extremely low illumination conditions underground. Background Technology

[0002] With the development of intelligent production and daily life, the application of intelligent monitoring is increasing. In underground working environments such as coal mines and tunnels, real-time perception and behavior recognition of personnel and equipment status are crucial. It can be used to monitor risk situations such as personnel location, abnormal stay, and equipment overheating, and provide timely warnings to ensure safe production. It is also suitable for complex scenarios such as emergency rescue and unmanned inspection.

[0003] Traditional downhole visual monitoring systems and methods suffer from numerous problems. For instance, in extremely low light conditions or even without visible light, the signal-to-noise ratio of images acquired by conventional visible light cameras drops significantly, resulting in a severe loss of detail. While single infrared thermal imaging can detect temperature distribution, it lacks texture and edge features, affecting target identification accuracy. Furthermore, existing dual-light imaging schemes mostly employ post-processing software fusion, failing to achieve deep collaboration at the structural, hardware, and algorithm levels. This leads to limited fusion efficiency, high power consumption, and difficulty in simultaneously meeting the comprehensive requirements of intrinsically safe explosion-proof operation, low power consumption, and high-reliability imaging in flammable and explosive environments such as methane. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a passive physical enhancement imaging method and device for dual-light fusion in extremely low-light environments underground. This method addresses the issues in underground environments such as coal mines and tunnels, where conventional visible light imaging equipment suffers from severely reduced image signal-to-noise ratios and loss of detail due to extremely low illumination. Single infrared thermal imaging lacks texture and edge features, affecting target identification accuracy. Existing dual-light imaging schemes fail to achieve deep fusion at the structural, hardware, and algorithmic levels, resulting in limited fusion efficiency, high power consumption, and difficulty in simultaneously meeting the comprehensive requirements of intrinsically safe explosion-proof operation, low power consumption, and high-reliability imaging.

[0005] The present invention provides a downhole extremely low illumination dual-light fusion passive physical enhancement imaging method, the imaging method comprising the following steps: S1, under the condition of no active supplementary lighting, a downhole low-light visual image is acquired by an electronic physical enhancement imaging module, and an infrared thermal distribution image in the 8–14 μm band is acquired by an uncooled infrared thermal imaging module. S2. Input the low-light visual image into the FPGA platform and perform real-time noise reduction and contrast enhancement processing based on the Poisson-Gaussian mixed noise model; input the infrared thermal distribution image into the correction unit at the back end of the CMOS readout circuit and perform non-uniformity correction based on the built-in blackbody reference source or scene adaptive algorithm. S3. Perform hierarchical fusion on the two preprocessed image data streams. The fusion level includes at least one of pixel level, feature level and decision level. S4. Input the fused image or feature data into the lightweight AI inference module to perform intelligent recognition of the behavior of personnel or equipment downhole, and output the recognition results and enhanced image stream through the gigabit Ethernet interface or fiber optic interface.

[0006] Preferably, step S1 specifically includes the following steps: S1.1 In the electronic physics-enhanced imaging module, the GaAs photocathode receives weak photon signals from the environment, with a spectral response range covering 400–950 nm and a quantum efficiency of not less than 30%. S1.2, Photoelectrons emitted by the photocathode enter the microchannel plate (MCP). Under the control of the automatic gating power supply, the operating voltage of the MCP is dynamically adjusted within the range of 600–1000V, achieving 10³–10 6 Electron gain; S1.3 The multiplied electron beam bombards the fluorescent screen to generate a visible light image, which is transmitted to the back-illuminated CMOS sensor via a fiber optic cone in a direct coupling manner. S1.4 In the uncooled infrared thermal imaging module, the vanadium oxide (VOx) focal plane array receives 8–14 μm infrared radiation. The resistance of each pixel changes with temperature. The CMOS readout circuit converts the resistance change into a voltage signal to form the original thermal image frame.

[0007] Preferably, step S2 specifically includes the following steps: S2.1 The FPGA receives the BT.656 format video stream output from the CMOS and uses a variational denoising model based on the Poisson-Gaussian joint distribution to perform spatial-temporal joint filtering on each frame of low-light image. S2.2 Perform local contrast stretching and gamma correction on the denoised image to enhance edge and texture details; S2.3 The infrared image correction unit triggers shutter-type blackbody reference calibration every 30 seconds, or runs a two-point correction algorithm based on scene statistical characteristics under no shutter conditions to compensate for the fixed-mode noise of the focal plane array. S2.4 The resolution of the corrected thermal image is aligned with that of the low-light image to the same coordinate system, with a pixel correspondence error not exceeding ±1 pixel.

[0008] Preferably, step S3 specifically comprises the following steps: S3.1 Deploy the dual-tree complex wavelet transform (DTCWT) kernel function in the FPGA to extract high-frequency subband coefficients from the low-light image and low-frequency approximation coefficients from the thermal image, and then synthesize the fused image by weighting according to the information entropy maximization criterion. S3.2. Run the ORB feature detector on the ARM processor to extract corner points and orientation-invariant features from the low-light image; simultaneously run the thermal contour segmentation algorithm to extract connected regions from the thermal image whose temperature is more than 3°C higher than the ambient average. S3.3. Perform spatial projection matching between ORB feature points and thermal anomaly regions. If the center distance between the two is less than a preset threshold, they are marked as credible target candidates. S3.4 Input the target classification results of the micro-light channel and the temperature attributes of the thermal channel into the lightweight Transformer model deployed on the NPU. The model outputs the final category label after confidence weighting.

[0009] Preferably, step S4 specifically includes the following steps: S4.1 The lightweight AI model adopts channel pruning and 8-bit integer quantization strategies, compressing the number of model parameters to less than 1 / 5 of the original model, and controlling the inference latency to within 20ms; S4.2 The recognition results include target category, location coordinates, temperature attribute and behavior status, and are encapsulated as a JSON format data packet; S4.3 Enhanced image stream and structured data are output via RJ45 Gigabit Ethernet port or ST single-mode fiber optic interface, supporting ONVIF protocol and GB / T 28181 standard; S4.4 The output signal is connected to the external monitoring network after being protected by an isolation transformer and a TVS diode array to prevent surge and electrostatic interference.

[0010] A downhole ultra-low illumination dual-light fusion passive physical enhancement imaging device, the imaging device comprising: A passive dual-light acquisition module is used to simultaneously acquire low-light visual images and infrared thermal distribution images in the absence of an active light source. The real-time preprocessing module is used to perform physical model-driven noise reduction and enhancement on low-light images and to perform non-uniformity correction on thermal images. The hierarchical fusion processing module is used to perform multi-level image fusion at the pixel level, feature level, and decision level. The intelligent recognition output module is used to identify target behavior based on fused data and output the results through a standard interface.

[0011] Preferably, the passive dual-light acquisition module specifically includes: The GaAs photocathode assembly has its photocathode surface facing the object side. It is aligned and installed with the microchannel plate through a vacuum-sealed glass shell, with the spacing controlled at 0.2±0.05mm. The microchannel board is connected to the ceramic substrate via indium pillar flip-chip bonding, and the substrate integrates an automatic gating high-voltage drive circuit. A fluorescent screen is deposited on the input end face of the fiber optic cone, and the output end face of the fiber optic cone is directly bonded to the photosensitive surface of the CMOS sensor via optical adhesive. The vanadium oxide focal plane array is integrated with the CMOS readout circuit via flip-chip bonding and is encapsulated in a TO-8 metal housing. The window is coated with an 8-14μm bandpass filter.

[0012] Preferably, the real-time preprocessing module specifically includes: The FPGA chip is a Xilinx Artix-7 XC7A100T, which internally implements a Poisson-Gaussian noise reduction IP core and a DTCWT transform core. The infrared calibration unit contains EEPROM to store calibration parameters, which are loaded into the bias DAC register of the CMOS readout circuit after each power-on. The image alignment unit uses a bilinear interpolation algorithm and constructs a lookup table in the FPGA to realize the mapping from thermal image pixels to low-light image coordinates.

[0013] Preferably, the layered fusion processing module specifically includes: The DTCWT fusion unit is configured with a 6-level decomposition depth, with the high-frequency subband using the L1 norm selection rule and the low-frequency subband using the average fusion rule. The feature matching unit runs the OpenCV 4.5 library on an ARM Cortex-A53 core, and the maximum number of ORB feature points is set to 500. The decision fusion unit is deployed on the Cambricon MLU220 NPU, running a MobileViT-S model optimized by TensorRT with an input size of 224×224×2 (dual channel).

[0014] Preferably, the intelligent recognition output module specifically includes: The AI ​​inference engine uses the INT8 quantization format, and the weights are stored in SPI NOR Flash and loaded into the NPU's on-chip SRAM at startup. The network interface controller integrates a PHY layer chip, supporting 10 / 100 / 1000BASE-T adaptive rates; The fiber optic interface module is packaged in an SFP package, has a built-in optoelectronic conversion chip, a transmission rate of 1.25Gbps, and a maximum transmission distance of 10km.

[0015] This invention provides a method and apparatus for downhole, extremely low-light dual-light fusion passive physical enhancement imaging. It includes at least the following beneficial effects: 1. This invention simultaneously acquires low-light visual images and infrared thermal distribution images under conditions without active supplemental lighting, and employs electrophysical enhancement and non-uniformity correction techniques to significantly improve the imaging signal-to-noise ratio and detail preservation capability in extremely low-light environments downhole, effectively solving the problems of poor image quality and difficulty in target identification in traditional visible light imaging under low-light conditions.

[0016] 2. This invention employs a hierarchical fusion strategy that combines pixel-level, feature-level, and decision-level approaches. It achieves hardware acceleration and algorithm optimization on a multi-core heterogeneous platform of FPGA, ARM, and NPU. This approach preserves the texture details of low-light images and the temperature information of infrared images while improving fusion efficiency and recognition accuracy. It overcomes the shortcomings of incomplete information in single imaging modes and poor real-time performance of software fusion.

[0017] 3. This invention adopts a lightweight AI inference model. Through channel pruning and 8-bit integer quantization technology, the number of model parameters is compressed to less than 1 / 5 of the original model, and the inference latency is controlled within 20ms. While ensuring the recognition accuracy, it achieves low power consumption and high real-time target behavior analysis, meeting the requirements of intrinsic safety explosion-proof and long-term stable operation in underground mines.

[0018] 4. The output interface of this invention is compatible with Gigabit Ethernet and fiber optic transmission, supports ONVIF protocol and GB / T 28181 standard, and is easy to connect to existing security monitoring systems; at the same time, it adopts isolation transformer and TVS diode array to provide electrical protection, which enhances the anti-interference ability and reliability of the system in the complex electromagnetic environment underground, and is suitable for unattended intelligent monitoring in high-risk environments such as coal mines and tunnels. Attached Figure Description

[0019] Figure 1 This is a block diagram of the overall architecture of the device of the present invention; Figure 2 This is a block diagram of the internal structure of the passive dual-light acquisition module in an embodiment of the present invention; Figure 3 This is a block diagram of the layered fusion processing module in an embodiment of the present invention; Figure 4 This is a block diagram of the intelligent recognition output module in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides a method and apparatus for passive physical enhancement imaging in extremely low-light downhole environments using dual-light fusion, the specific implementation of which is described below. (Combined with...) Figures 1 to 4 The structure and process shown below will be described in detail below, including the various components, connections, signal flow, processing logic, and operation process of the present invention, to ensure that those skilled in the art can fully implement the technical solution.

[0022] like Figure 1 As shown, this invention consists of four core modules: a passive dual-light acquisition module, a real-time preprocessing module, a hierarchical fusion processing module, and an intelligent recognition output module. These modules are sequentially connected via board-level high-speed signal lines, parallel video buses, and on-chip interconnect buses, forming a complete image acquisition-preprocessing-fusion-recognition-output link. Specifically, the passive dual-light acquisition module simultaneously acquires low-light visual images and infrared thermal distribution images through optical and photoelectric conversion mechanisms, and outputs them to the real-time preprocessing module. The real-time preprocessing module performs noise reduction, enhancement, and correction on both images, and then sends the processing results to the hierarchical fusion processing module. The hierarchical fusion processing module executes a multi-level fusion strategy to generate a fused image or structured feature data. Finally, the intelligent recognition output module performs target behavior recognition based on the fusion results and outputs the image stream and structured data through a standard communication interface.

[0023] Further reference Figure 2The passive dual-light acquisition module internally comprises a low-light imaging subsystem and an infrared imaging subsystem, which share the same optical window but employ independent optical path designs. The low-light imaging subsystem consists of a GaAs photocathode, a microchannel plate, a fluorescent screen, a fiber optic cone, and a back-illuminated CMOS sensor arranged sequentially. The GaAs photocathode is located at the front end of the vacuum-sealed glass housing, with its photocathode surface facing the object, receiving photons from the weak visible to near-infrared wavelengths (400-950nm). After photoelectrons are emitted from the photocathode, they enter the microchannel plate under the influence of an electric field. The microchannel plate is fixed to a ceramic substrate using an indium pillar flip-chip bonding process. This substrate integrates an automatically gated high-voltage drive circuit, which can dynamically adjust the MCP operating voltage within the range of 600-1000V. The multiplied electron beam bombards the fluorescent screen deposited on the input end face of the fiber optic cone, exciting a visible light image. This image is transmitted to the photosensitive surface of the CMOS sensor via direct coupling through the fiber optic cone. The two are bonded together with optical adhesive to ensure that there are no air gaps in the optical path, reducing light loss and aberrations. The infrared imaging subsystem consists of a vanadium oxide focal plane array, a CMOS readout circuit, and a bandpass filter window. The vanadium oxide focal plane array is integrated with the CMOS readout circuit via flip-chip bonding and is encapsulated in a TO-8 metal housing. The front window of the housing is coated with an 8-14μm bandpass filter film to allow only infrared radiation of that band to pass through. The resistance of the vanadium oxide material in each pixel changes with temperature. The CMOS readout circuit converts the resistance change into a voltage signal through a column-parallel readout architecture to form the original thermal image frame. The micro-photonics subsystem and the infrared subsystem are arranged side by side in physical layout and share the same mechanical housing, but their optical paths are isolated from each other to avoid crosstalk.

[0024] The real-time preprocessing module receives BT.656 format video streams from the CMOS sensor and raw thermal image data from the CMOS readout circuit. This module uses a Xilinx Artix-7 XC7A100T FPGA chip as its core, internally implementing a Poisson-Gaussian denoising IP core and a dual-tree complex wavelet transform (DTCWT) core. The FPGA is connected to the output pins of the CMOS sensor via an 8-bit parallel data bus, synchronously receiving pixel clock, line synchronization, and field synchronization signals. For each frame of low-light image, the FPGA first executes a variational denoising model based on a Poisson-Gaussian joint distribution. This model uses a 3×3 neighborhood window in the spatial domain and constructs a four-dimensional joint filter using the buffered data from the previous frame in the temporal domain, completing the spatial-temporal joint filtering. Subsequently, local contrast stretching is performed on the denoised image, and the local mean and standard deviation are calculated using a sliding window, according to the formula... Normalization enhancement is performed, where α and β are configurable parameters; finally, gamma correction is performed, with the γ value set to 0.6 to improve details in dark areas. For infrared images, an infrared correction unit is integrated at the back end of the CMOS readout circuit. This correction unit contains an EEPROM that stores the bias and gain parameters calibrated at the factory. After each power-on, the contents of the EEPROM are loaded into the bias DAC register inside the CMOS readout circuit. During operation, a shutter-type blackbody reference calibration is triggered every 30 seconds, which controls a miniature shutter to block the infrared field of view, so that the vanadium oxide focal plane array only receives the built-in blackbody radiation, and the correction parameters are updated accordingly. If the device is not equipped with a shutter, a two-point correction algorithm based on scene statistical characteristics is run, which uses the bimodal characteristics of the temperature distribution in the image to estimate the high and low temperature reference points and compensate for fixed-mode noise. The corrected thermal image needs to be aligned with the low-light image. The image alignment unit builds a lookup table in the FPGA. This lookup table is obtained through calibration experiments and records the mapping position of each pixel of the thermal image in the low-light image coordinate system. A bilinear interpolation algorithm is used to achieve sub-pixel level alignment to ensure that the pixel correspondence error does not exceed ±1 pixel.

[0025] Layered fusion processing module, such as Figure 3As shown, the system includes a DTCWT pixel-level fusion unit, an ORB and thermal contour feature extraction unit, a spatial matching unit, and a lightweight Transformer decision fusion unit. The DTCWT pixel-level fusion unit is deployed inside the FPGA. It performs a 6-level dual-tree complex wavelet decomposition on the low-light image input to extract high-frequency sub-band coefficients (including horizontal, vertical, and diagonal details). It also performs a 6-level decomposition on the thermal image input to extract low-frequency approximation coefficients. The high-frequency sub-band uses an L1 norm selection rule, calculating the L1 norm for the high-frequency coefficients of both the low-light and thermal images at the same location and retaining the one with the larger norm. The low-frequency sub-band uses an average fusion rule, taking the arithmetic mean of the two. After fusion, a pixel-level fused image is generated through inverse DTCWT reconstruction. The ORB and thermal contour feature extraction unit run on an ARM Cortex-A53 quad-core processor, loaded with the OpenCV 4.5 library. Corner features are extracted from the low-light image using the ORB detector, with a maximum of 500 feature points. Each feature point includes location, scale, orientation, and a BRIEF descriptor. Simultaneously, the thermal image is processed by a thermal contour segmentation algorithm. First, the average temperature T_mean of the entire image is calculated. Then, all pixels with temperatures higher than T_mean+3℃ are marked. Closed regions are extracted through connected component analysis to form a set of thermal anomaly regions. The spatial matching unit performs spatial projection matching between the ORB feature point set and the thermal anomaly region set, calculating the Euclidean distance from each ORB feature point to the centroid of each thermal anomaly region. If the minimum distance is less than a preset threshold (e.g., 10 pixels), the feature point is associated with the corresponding thermal region and marked as a reliable target candidate. The lightweight Transformer decision fusion unit is deployed on the Cambricon MLU220 NPU. It receives the target classification results from the low-light channel (preliminary classification by a lightweight CNN) and the temperature attributes (such as the average temperature of the region and the temperature rise) from the thermal channel. The input is a two-channel tensor with a size of 224×224×2, where the first channel is the cropped region of the low-light image and the second channel is the normalized temperature map of the corresponding thermal image region. The model adopts the MobileViT-S architecture and supports INT8 inference after TensorRT optimization. It outputs the final category label (such as "personnel", "equipment", "obstacle") after confidence weighting.

[0026] Intelligent recognition output module such as Figure 4As shown, the system consists of an NPU inference unit, SPI NOR Flash, a Gigabit Ethernet PHY chip, an RJ45 interface, an SFP fiber optic module, and isolation protection circuitry. The NPU inference unit is a Cambricon MLU220 chip with 16MB of on-chip SRAM used to store runtime model weights and intermediate feature maps. The model weights are stored in INT8 quantization format in an external SPI NOR Flash of 128MB, loaded into the NPU's on-chip SRAM via DMA through the QSPI interface during system startup. The inference latency is controlled to within 20ms. The recognition results include the target category, bounding box coordinates (x, y, w, h), average temperature value, and behavioral state (e.g., "stationary," "moving," "fallen"), encapsulated as a JSON format data packet. The enhanced image stream and structured data are output via two parallel paths: one path connects to an RJ45 interface via a Gigabit Ethernet PHY chip (PHY chip 33 supports 10 / 100 / 1000BASE-T adaptive rates and complies with the IEEE 802.3 standard); the other path outputs via an SFP fiber optic module. This module uses a standard SFP package, incorporates a photoelectric conversion chip, supports a transmission rate of 1.25Gbps, and has a maximum transmission distance of 10km, suitable for long-distance underground communication. Both output signals undergo isolation protection circuitry before connecting to the external network. This circuitry includes an isolation transformer and a TVS diode array. The isolation transformer provides electrical isolation, and the TVS diode array clamps surge voltages to prevent electrostatic discharge (ESD) and lightning surges from damaging the internal circuitry. The entire device is powered by an intrinsically safe power supply, operating at 12V DC, with a total power consumption of less than 8W, meeting the intrinsic safety and explosion-proof requirements for underground coal mines.

[0027] In practical underground applications, the device is installed on the sidewall or roof support of the tunnel, covering the critical work area. Without any active lighting, the GaAs photocathode continuously receives natural scattered light or weak light sources such as equipment indicator lights within the tunnel. The microchannel plate automatically adjusts its gain according to the ambient brightness to ensure stable output image brightness. Simultaneously, the vanadium oxide focal plane array senses the infrared radiation emitted by personnel or equipment, forming a thermal distribution map. After preprocessing and fusion, the two images clearly present the outline of personnel and their surface texture, effectively identifying them even in complete darkness or smoky environments. The AI ​​inference module determines in real time whether abnormal behavior exists (such as prolonged inactivity of personnel or overheating of equipment) and uploads the alarm information along with the enhanced image to the ground monitoring center via fiber optic link. It supports the ONVIF protocol and the GB / T 28181 standard, facilitating seamless integration with existing security platforms. The entire system requires no active light source, has no moving parts, is compact, and highly reliable, making it suitable for long-term unattended monitoring in high-risk, low-light environments such as coal mines and tunnels.

[0028] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for passive physical enhancement imaging in extremely low-light downhole environments using dual-light fusion, characterized in that, Includes the following steps: S1. Under conditions without active supplemental lighting, the downhole low-light visual image is acquired through the electronic physical enhancement imaging module, and the infrared thermal distribution image in the 8-14μm band is acquired through the uncooled infrared thermal imaging module. S2. Input the low-light visual image into the FPGA platform and perform real-time noise reduction and contrast enhancement processing based on the Poisson-Gaussian mixed noise model; input the infrared thermal distribution image into the correction unit at the back end of the CMOS readout circuit and perform non-uniformity correction based on the built-in blackbody reference source or scene adaptive algorithm. S3. Perform hierarchical fusion on the two preprocessed image data streams. The fusion level includes at least one of pixel level, feature level and decision level. S4. Input the fused image or feature data into the lightweight AI inference module to perform intelligent recognition of the behavior of personnel or equipment downhole, and output the recognition results and enhanced image stream through the gigabit Ethernet interface or fiber optic interface.

2. The downhole ultra-low illumination dual-light fusion passive physical enhancement imaging method according to claim 1, characterized in that, Step S1 specifically includes the following steps: S1.1 In the electronic physics-enhanced imaging module, the GaAs photocathode receives weak photon signals from the environment, with a spectral response range covering 400–950 nm and a quantum efficiency of not less than 30%. S1.2, Photoelectrons emitted by the photocathode enter the microchannel plate. Under the control of the automatic gating power supply, the operating voltage of the microchannel plate is dynamically adjusted within the range of 600-1000V, achieving 10³-10 6 Electron gain; S1.3 The multiplied electron beam bombards the fluorescent screen to generate a visible light image, which is transmitted to the back-illuminated CMOS sensor via a fiber optic cone in a direct coupling manner. S1.4 In the uncooled infrared thermal imaging module, the vanadium oxide focal plane array receives 8-14μm infrared radiation. The resistance of each pixel changes with temperature. The CMOS readout circuit converts the resistance change into a voltage signal to form the original thermal image frame.

3. The downhole ultra-low illumination dual-light fusion passive physical enhancement imaging method according to claim 1, characterized in that, Step S2 specifically includes the following steps: S2.1 The FPGA receives the BT.656 format video stream output by the CMOS sensor and uses a variational denoising model based on the Poisson-Gauss joint distribution to perform spatial-temporal joint filtering on each frame of low-light image. S2.2 Perform local contrast stretching and gamma correction on the denoised image to enhance edge and texture details; S2.3 The infrared image correction unit triggers shutter-type blackbody reference calibration every 30 seconds, or runs a two-point correction algorithm based on scene statistical characteristics under no shutter conditions to compensate for the fixed-mode noise of the focal plane array. S2.4 The resolution of the corrected thermal image is aligned with that of the low-light image to the same coordinate system, with a pixel correspondence error not exceeding ±1 pixel.

4. The downhole ultra-low illumination dual-light fusion passive physical enhancement imaging method according to claim 1, characterized in that, Step S3 specifically includes the following steps: S3.1 Deploy a dual-tree complex wavelet transform kernel function in the FPGA to extract high-frequency subband coefficients from the low-light image and low-frequency approximation coefficients from the thermal image, and then synthesize the fused image by weighting according to the information entropy maximization criterion. S3.

2. Run the ORB feature detector on the ARM processor to extract corner points and orientation-invariant features from the low-light image; simultaneously run the thermal contour segmentation algorithm to extract connected regions from the thermal image whose temperature is more than 3°C higher than the ambient average. S3.

3. Perform spatial projection matching between ORB feature points and thermal anomaly regions. If the center distance between the two is less than a preset threshold, they are marked as credible target candidates. S3.4 Input the target classification results of the micro-light channel and the temperature attributes of the thermal channel into the lightweight Transformer model deployed on the NPU. The model outputs the final category label after confidence weighting.

5. The downhole ultra-low illumination dual-light fusion passive physical enhancement imaging method according to claim 1, characterized in that, Step S4 specifically includes the following steps: S4.1 The lightweight AI model adopts channel pruning and 8-bit integer quantization strategies, compressing the number of model parameters to less than 1 / 5 of the original model, and controlling the inference latency to within 20ms; S4.2 The recognition results include target category, location coordinates, temperature attribute and behavior status, and are encapsulated as a JSON format data packet; S4.3 Enhanced image stream and structured data are output via RJ45 Gigabit Ethernet port or ST single-mode fiber optic interface, supporting ONVIF protocol and GB / T 28181 standard; S4.4 The output signal is connected to the external monitoring network after being protected by an isolation transformer and a TVS diode array to prevent surge and electrostatic interference.

6. A downhole ultra-low illumination dual-light fusion passive physical enhancement imaging device, characterized in that, include: A passive dual-light acquisition module is used to simultaneously acquire low-light visual images and infrared thermal distribution images in the absence of an active light source. The real-time preprocessing module is used to perform physical model-driven noise reduction and enhancement on low-light images and to perform non-uniformity correction on thermal images. The hierarchical fusion processing module is used to perform multi-level image fusion at the pixel level, feature level, and decision level. The intelligent recognition output module is used to identify target behavior based on fused data and output the results through a standard interface.

7. The downhole ultra-low illumination dual-light fusion passive physical enhancement imaging device according to claim 6, characterized in that, The passive dual-light acquisition module specifically includes: The GaAs photocathode has its photocathode surface facing the object side and is aligned and installed with the microchannel plate through a vacuum-sealed glass shell with a spacing of 0.2 mm. The microchannel board is connected to the ceramic substrate via indium pillar flip-chip bonding, and the substrate integrates an automatic gating high-voltage drive circuit. A fluorescent screen is deposited on the input end face of a fiber optic taper, and the output end face of the fiber optic taper is directly bonded to the photosensitive surface of a CMOS sensor using optical adhesive. The vanadium oxide focal plane array is integrated with the CMOS readout circuit via flip-chip bonding and is encapsulated in a TO-8 metal housing. The window is coated with an 8-14μm bandpass filter as a bandpass filter window.

8. The downhole ultra-low illumination dual-light fusion passive physical enhancement imaging device according to claim 6, characterized in that, The real-time preprocessing module specifically includes: The FPGA chip is a Xilinx Artix-7 XC7A100T, which internally implements a Poisson-Gaussian denoising IP core and a dual-tree complex wavelet transform core. The infrared calibration unit contains EEPROM to store calibration parameters, which are loaded into the bias DAC register of the CMOS readout circuit after each power-on. The image alignment unit uses a bilinear interpolation algorithm and constructs a lookup table in the FPGA to realize the mapping from thermal image pixels to low-light image coordinates.

9. A downhole ultra-low illumination dual-light fusion passive physical enhancement imaging device according to claim 6, characterized in that, The layered fusion processing module specifically includes: The DTCWT fusion unit is configured with a 6-level decomposition depth, with the high-frequency subband using the L1 norm selection rule and the low-frequency subband using the average fusion rule. The feature matching unit runs the OpenCV 4.5 library on an ARM Cortex-A53 core, with a maximum of 500 ORB feature points. The decision fusion unit is deployed on the Cambricon MLU220 NPU, running a MobileViT-S model optimized with TensorRT, with an input size of 224×224×2.

10. A downhole ultra-low illumination dual-light fusion passive physical enhancement imaging device according to claim 6, characterized in that, The intelligent recognition output module specifically includes: The AI ​​inference engine uses the INT8 quantization format, and the weights are stored in SPI NOR Flash and loaded into the NPU's on-chip SRAM at startup. The network interface controller integrates a PHY layer chip, supporting 10 / 100 / 1000BASE-T adaptive rates; The fiber optic interface module is packaged in an SFP package, has a built-in photoelectric conversion chip, a transmission rate of 1.25Gbps, and a maximum transmission distance of 10km.

Images