An intelligent inspection system based on hyperspectral and evidence image atlas

Abstract

The present application relates to a kind of based on super wide spectrum and evidence image atlas integration intelligent inspection system, belong to intelligent trace detection technical field.Wherein, the system includes: through the continuous spectral image of evidence in 400nm-1000nm wave band by hyperspectral image acquisition unit acquisition, based on spectral camera completes spectral data acquisition, by the fusion output evidence atlas integration data set of preprocessed data;Through short-wave infrared hyperspectral acquisition unit acquisition evidence in 1000nm-2500nm wave band hyperspectral data, based on motor-driven interferometer moving mirror movement, different wavelength short-wave infrared light interference occurs and forms interference chart, interference chart is handled and exports evidence short-wave infrared atlas;Through full-spectrum image acquisition unit acquisition evidence in 20nm-2500nm wave band wide spectrum image, CMOS sensor receives the full wave band light reflected by evidence, capture color information and the infrared reflection difference of material, generate evidence full-spectrum image after processing.It realizes the high sensitivity, high accuracy, high efficiency, high traceability of evidence inspection.

Description

Technical Field

[0001] This invention belongs to the field of intelligent trace evidence detection technology, specifically relating to an intelligent inspection system that integrates ultrawide spectrum and physical evidence images. Background Technology

[0002] Traditional forensic examination relies on a separation between visible light imaging and single-spectral analysis. Visible light imaging is limited to the 400-760nm wavelength range, only capturing the morphology and color information of material surfaces. In actual cases, many key forensic features lie in the invisible spectral region: for example, sweat fingerprints show fluorescence under 254nm ultraviolet light excitation, but ordinary cameras cannot record them; the near-infrared band can penetrate the ink on the surface of paper, clearly revealing the original handwriting covered by it; and short-wave infrared can accurately distinguish chemically similar cotton and linen blends based on the differences in fiber absorption of specific wavelengths. Statistics show that some trace evidence is beyond the visible light detection range, making feature extraction difficult and seriously affecting the efficiency of case solving.

[0003] In traditional testing procedures, spectrometers need to collect reflectance and absorption spectral data of samples point by point to generate wavelength-intensity curves reflecting the composition of substances; while the morphological information of physical evidence is obtained separately by a camera. This separate testing mode has three drawbacks: First, data integration relies on manual comparison, and the average time from data acquisition to result analysis for a single sample is too long; second, inspectors need to rely on experience to match spectral curves with spatial locations, and the judgment results of different experts may differ, leading to disputes over the chain of evidence in complex cases; third, the single-point sampling method of spectrometers cannot reflect the spatial distribution characteristics of physical evidence, making it difficult to achieve complete analysis for samples with gradually changing components.

[0004] In real-world testing scenarios, complex backgrounds and ambient light interference pose significant challenges to traditional techniques. For example, light-colored fibers on dark fabrics are easily obscured by the background under visible light due to their low color contrast; reflections from metal surfaces can cause localized overexposure, resulting in the loss of potential fingerprint features; and the spectral superposition effect under multi-light source environments can severely interfere with the accurate identification of material components. Traditional techniques lack multi-band information collaborative processing mechanisms, leading to a high rate of false negatives and a significantly increased risk of misjudgment when faced with such interference, severely impacting the reliability of test results. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides an intelligent inspection system integrating ultrawide spectrum and physical evidence images. The objective of this invention can be achieved through the following technical solutions:

[0006] A unified intelligent examination system based on ultrawide spectrum and physical evidence images, comprising:

[0007] The hyperspectral image acquisition unit is used to acquire continuous spectral images of the evidence in the 400nm-1000nm band. It sends the working band requirement to the multispectral light source control module via a communication interface, triggering the activation of the visible and near-infrared light source. The multispectral light source control module presets the initial illuminance based on the material of the evidence. The pushbroom-type spectral camera adopts a linear pushbroom mode, using an internal scanning mechanism to move the spectral sensor along the surface of the evidence to complete spectral data acquisition. The acquired data is preprocessed based on dark current correction, spectral calibration, and bad pixel repair techniques. The preprocessed data is then fused by the image fusion processing module to output a unified evidence image-spectral dataset.

[0008] The short-wave infrared hyperspectral acquisition unit is used to acquire hyperspectral data of physical evidence in the 1000nm-2500nm band. The short-wave infrared hyperspectral analyzer sends a short-wave infrared band request to the multi-band light source system via a communication protocol, triggering the halogen tungsten lamp light source to start. The optical path calibration module drives the built-in standard reference board into the optical path to acquire the reference spectrum. The interferometer's moving mirror moves under the drive of a motor, generating a change in optical path difference, causing short-wave infrared light of different wavelengths to interfere and form an interferogram. The interferogram is corrected, wavelength calibrated, and reflectivity converted to output the short-wave infrared spectrum of the physical evidence.

[0009] A full-spectrum image acquisition unit is used to acquire broadband images of the evidence in the 20nm-2500nm wavelength range; a full-spectrum image analyzer sends a full-spectrum wavelength request to the multi-band light source system, triggering the multi-spectral LED array light source to start and configuring the initial color temperature and initial illuminance; an optical path calibration module drives a standard color chart into the field of view to acquire calibration images; the ultraviolet enhancement lens and anti-reflection coating in the broadband imaging component work together to receive the full-spectrum light reflected by the evidence through a CMOS sensor, capturing color information and differences in infrared reflection of the material; noise removal, color restoration, and band fusion processing are performed on the received data to generate a full-spectrum image of the evidence.

[0010] As a preferred embodiment of the present invention, the activation method of the visible and near-infrared light source in the hyperspectral image acquisition unit is as follows:

[0011] The hyperspectral image acquisition unit determines the target wavelength range, light intensity requirements, color temperature parameters, and core operating parameters according to the preset acquisition task. Based on the requirements of hyperspectral analysis, the target wavelength range of 400nm-760nm is used to capture the color and basic material characteristics of physical evidence, and the target wavelength range of 760nm-1000nm is used to detect the near-infrared absorption characteristics of material molecules.

[0012] The communication control submodule of the hyperspectral image acquisition unit encapsulates the core operating parameters into instruction frames according to a preset protocol format, converts the TTL signal into an RS485 differential signal based on a level conversion circuit, and sends the instruction frames to the communication receiving end of the multispectral light source control module through a shielded twisted pair cable to achieve data transmission. Based on the instruction frames received by the communication receiving end, it performs identification verification, parses the parameter area data, and extracts the core operating parameters. Based on the core operating parameters, it sends a control signal to the driving circuit of the visible and near-infrared light source to start the light source.

[0013] Specifically, the method for acquiring spectral data in the hyperspectral image acquisition unit is as follows:

[0014] Based on the inspection requirements, the spectral band range is locked. The hyperspectral image acquisition unit automatically transmits the band parameters to the tunable filter group of the pushbroom spectrophotometer. The filters in the tunable filter group complete the initial positioning according to the wavelength requirements. The motorized zoom lens automatically adjusts the focal length according to the size of the evidence and confirms the image clarity through the laser focusing assistance system.

[0015] The linear array sensor of the pushbroom spectrophotometer is aligned with the narrow strip area of ​​the evidence. The incident light is focused by the lens and enters the spectrometer, where it is decomposed into a continuous spectrum of 400nm-1000nm. Light signals of different wavelengths are focused onto different pixel positions of the linear array sensor. After a single exposure, a line of raw spectral data containing 200 bands is generated.

[0016] The scanning drive mechanism drives the camera to move along a direction perpendicular to the linear array. The movement step is matched with the pixel size, and the spectral data of the evidence is collected line by line. The collected spectral data is stitched together in the scanning order to generate a three-dimensional data structure containing spatial and spectral dimensions.

[0017] Specifically, the method for fusing preprocessed data in the hyperspectral image acquisition unit is as follows:

[0018] A spectral feature extraction network comprising convolutional layers, pooling layers, and residual connections is constructed. The convolutional layers capture multi-scale spectral information of evidence from local details to global structure. The pooling layers reduce data dimensionality while preserving key features. The residual connections enhance the learning ability of the spectral feature extraction network for complex spectral features. A spatial feature extraction network captures multi-scale features of evidence images from microscopic texture details to macroscopic geometric contours.

[0019] By calculating the mutual information matrix of spectral features and spatial features, relevant feature pairs are identified, and the correlation between spectral features and spatial features is established based on the relevant feature pairs. The spectral feature vector and spatial feature vector are multiplied by the corresponding attention weights through an adaptive weighted fusion algorithm to obtain weighted spectral features. The two types of weighted features are fused into a joint feature vector through feature splicing and residual connection to generate a complete physical evidence image-spectrum integrated dataset.

[0020] Specifically, the feature splicing and residual connection are implemented as follows:

[0021] The spectral feature vector magnitude is unified by normalization, and the spatial feature vector value is compressed to a preset range by standardization. The hyperspectral image acquisition unit automatically verifies the consistency of the sample quantity of the two types of features. Based on the verification result, the standardized spectral feature vector and the spatial feature vector are directly concatenated along the channel dimension to generate an initial fused feature vector. The initial fused feature vector is input into the residual module, and joint features are formed by standardizing the mean and variance of the feature vector and compressing the dimension. The joint feature vector is associated with the data acquired by the hyperspectral image acquisition unit to output a complete dataset.

[0022] Specifically, the interferogram in the short-wave infrared hyperspectral acquisition unit is formed as follows:

[0023] The short-wave infrared hyperspectral analyzer performs optical path calibration, resets the interferometer moving mirror to the zero optical path difference position, and the laser collimation system emits an auxiliary positioning laser. By monitoring the position deviation of the reflected light spot, the tilt angle of the fixed mirror is adjusted to complete the convergence of the two beams on the target surface of the MCT detector.

[0024] After spectral acquisition is initiated, the servo motor drives the moving mirror of the interferometer to move along the optical axis. Based on the linear change in the reflected optical path difference between the fixed mirror and the moving mirror as the moving mirror moves, the two coherent beams converge at the MCT detector, and interference occurs due to the difference in optical path difference. The MCT detector converts the interference light intensity signal into an analog electrical signal, amplifies it through a preamplifier, and converts it into a digital signal by an analog-to-digital converter. The digital signal is then processed using dark current subtraction, bad pixel repair, and phase correction techniques to output the interferogram.

[0025] Specifically, after the short-wave infrared hyperspectral acquisition unit is started, it performs self-checks on the Fourier transform spectral component, MCT detector, optical adjustment mechanism, and data transmission link. Any abnormalities are displayed in real time through the terminal interaction module. The ultra-low temperature MCT detection module starts a precision temperature control program. Based on proportional-integral-derivative closed-loop control technology, the semiconductor cooler starts a graded cooling mode, which rapidly reduces the temperature of the MCT detector through high-power cooling. When the temperature approaches the target value, it switches to a low-power fine adjustment mode, which works in conjunction with the temperature sensor for real-time monitoring and feedback.

[0026] Specifically, the reflectance conversion in the short-wave infrared hyperspectral acquisition unit is implemented as follows:

[0027] The sample is placed on the stage, and the short-wave infrared hyperspectral analyzer scans it, causing the moving mirror of the interferometer to move and generate an optical path difference. The MCT detector receives the interference signal of the light reflected from the sample and generates the original light intensity spectrum based on the fast Fourier transform. During the optical path calibration stage, the light intensity data reflected by the standard reference plate is collected as the reference plate reference spectrum.

[0028] The original light intensity spectrum is subjected to dark current subtraction processing. Based on the reflectance of the reference plate's baseline spectrum and the processed original light intensity spectrum, a wavelength-specific calibration coefficient is calculated. The calculated calibration coefficient is then smoothed. Based on the calibration coefficient, the relative light intensity is converted to absolute reflectance by performing reflectance conversion on the channels of the original light intensity spectrum.

[0029] Specifically, the full-spectrum image acquisition unit supports pseudo-color synthesis, and the synthesis method is as follows:

[0030] Three bands were randomly selected within the 200nm-2500nm range. The reflectance values ​​of each band were linearly stretched to a grayscale value range of 0-255 to perform radiometric normalization on the selected bands. Median filtering was used to eliminate high-frequency noise and preserve edge features. The selected band images were spatially registered using a sub-pixel level registration algorithm.

[0031] Based on the band matching algorithm, the preprocessed selected band data is mapped to the red, green, and blue channels respectively. An adaptive white balance algorithm is used to monitor the spectral distribution characteristics of each pixel in the image in real time, and automatically adjust the color temperature and hue parameters based on the detection results. A combination of piecewise linear stretching and histogram equalization is used to remap the grayscale range of each channel to suppress color overflow. A local color enhancement module analyzes the texture and spectral characteristics of the trace area to enhance color contrast and edge sharpness. Based on the identification of the trace target and background area, the hue and saturation of the target area are optimized using a color mapping table to generate a pseudo-color image.

[0032] Specifically, the steps for the ultraviolet-enhanced lens and the anti-reflective coating to work together in the full-spectrum image acquisition unit are as follows:

[0033] Full-band light carrying physical evidence information enters the imaging component, and the ultraviolet enhancement lens based on the quartz glass substrate controls the attenuation of ultraviolet light; based on the optical interference structure, the outer film system of the anti-reflection coating suppresses the reflection of light in each band.

[0034] Based on low-absorption special optical materials, the ultraviolet-enhanced lens reduces ultraviolet light propagation loss at the substrate level; the anti-reflective coating achieves deep suppression of interface reflection through a multi-layer gradient refractive index ultraviolet-specific film system design; the aspherical geometry of the ultraviolet-enhanced lens and the refractive index modulation of the anti-reflective coating form an optical compensation mechanism, which corrects aberrations and focusing deviations in the ultraviolet band based on matching the refractive characteristics of ultraviolet and visible light, achieving sub-pixel-level spatial alignment between ultraviolet imaging and visible light imaging.

[0035] The beneficial effects of this invention are as follows:

[0036] The system utilizes three acquisition units to collaboratively cover an ultra-wide spectral range of 200nm-2500nm, overcoming the limitations of traditional single-band inspection. The hyperspectral imaging unit (400nm-1000nm) precisely captures visible light color features and near-infrared molecular absorption characteristics, identifying the spectral fingerprints of biological evidence such as dye components and bloodstains. The short-wave infrared hyperspectral acquisition unit (1000nm-2500nm) uses molecular vibrational spectral analysis to distinguish hidden features such as organic residues and fiber materials, filling the information blind spots in the visible light band. The full-spectrum imaging unit (200nm-2500nm) integrates ultraviolet fluorescence signals and broad-spectrum morphological information, revealing features invisible to the naked eye, such as sweat fingerprints and alteration marks. The combination of these three units forms a full-band feature network of "ultraviolet-visible-near-infrared-short-wave infrared," ensuring comprehensive and thorough examination of evidence.

[0037] The system employs multi-level preprocessing techniques to eliminate interference, transforming data from "raw signals" to "precise information." Dark current correction, bad pixel repair, and median filtering significantly improve the spectral signal-to-noise ratio, reducing image noise and preventing noise-induced feature identification. Spectral calibration and reflectance conversion transform relative light intensity data into absolute reflectance, enabling data comparability across different times and devices. Sub-pixel-level registration algorithms ensure spatial alignment of multi-band images, laying the foundation for fusion analysis. These technologies enable the system to achieve micrometer-level spatial resolution and nanometer-level spectral resolution, meeting the precise analysis requirements of trace evidence.

[0038] Through innovative algorithms in the spectral fusion processing module, the bottleneck of the separation between spectral and image information is overcome. The spectral feature network captures details of material composition, the spatial feature network extracts texture contours, and the mutual information matrix identifies the "composition-morphology" correlation. Attention weights dynamically allocate weights to important features, and feature splicing and residual connections preserve key information. The generated joint feature vector simultaneously contains the spectral fingerprint and spatial distribution patterns of the material. Pseudo-color synthesis technology transforms multi-band information into intuitive images, and band mapping and color enhancement visualize hidden features. The fused data enables "visible morphology and clear composition," significantly improving the efficiency of interpreting complex evidence.

[0039] The system reduces manual intervention through end-to-end intelligent control, ensuring long-term operational reliability. It automatically adjusts the wavelength range, exposure time, and light source parameters based on the type of evidence, reducing manual debugging costs. Optical path self-testing and temperature closed-loop control, along with regular standard board calibration, ensure long-term equipment stability. Hardware self-testing and data quality verification provide real-time feedback on anomalies, preventing the generation of invalid data. These designs significantly shorten the testing time for a single sample, while reducing human error and improving testing consistency.

[0040] The system expands its application boundaries by flexibly adapting to complex inspection scenarios. It achieves precise response for different materials such as fibers, paper, coatings, and biological traces through light source adaptation and spectral range adjustment. The combination of push-broom scanning and single-point spectral analysis meets the inspection needs from large-area evidence to tiny traces. It supports custom band combinations, feature extraction algorithms, and pseudo-color mapping tables, and can quickly expand its functions according to new inspection standards or special evidence types.

[0041] In conclusion, this invention can be widely applied in fields such as criminal investigation, cultural relic appraisal, and material analysis, providing comprehensive technical support for the scientific examination of complex physical evidence. Attached Figure Description

[0042] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0043] Figure 1 This is a schematic diagram of the process of an intelligent inspection system based on ultrawide spectrum and physical evidence image integration according to the present invention. Detailed Implementation

[0044] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0045] Please see Figure 1 A unified intelligent examination system based on ultrawide spectrum and physical evidence images, comprising:

[0046] The hyperspectral image acquisition unit is used to acquire continuous spectral images of the evidence in the 400nm-1000nm band. It sends the working band requirement to the multispectral light source control module via a communication interface, triggering the activation of the visible and near-infrared light source. The multispectral light source control module presets the initial illuminance based on the material of the evidence. The pushbroom-type spectral camera adopts a linear pushbroom mode, using an internal scanning mechanism to move the spectral sensor along the surface of the evidence to complete spectral data acquisition. The acquired data is preprocessed based on dark current correction, spectral calibration, and bad pixel repair techniques. The preprocessed data is then fused by the image fusion processing module to output a unified evidence image-spectral dataset.

[0047] The short-wave infrared hyperspectral acquisition unit is used to acquire hyperspectral data of physical evidence in the 1000nm-2500nm band. The short-wave infrared hyperspectral analyzer sends a short-wave infrared band request to the multi-band light source system via a communication protocol, triggering the halogen tungsten lamp light source to start. The optical path calibration module drives the built-in standard reference board into the optical path to acquire the reference spectrum. The interferometer's moving mirror moves under the drive of a motor, generating a change in optical path difference, causing short-wave infrared light of different wavelengths to interfere and form an interferogram. The interferogram is corrected, wavelength calibrated, and reflectivity converted to output the short-wave infrared spectrum of the physical evidence.

[0048] A full-spectrum image acquisition unit is used to acquire broadband images of the evidence in the 20nm-2500nm wavelength range; a full-spectrum image analyzer sends a full-spectrum wavelength request to the multi-band light source system, triggering the multi-spectral LED array light source to start and configuring the initial color temperature and initial illuminance; an optical path calibration module drives a standard color chart into the field of view to acquire calibration images; the ultraviolet enhancement lens and anti-reflection coating in the broadband imaging component work together to receive the full-spectrum light reflected by the evidence through a CMOS sensor, capturing color information and differences in infrared reflection of the material; noise removal, color restoration, and band fusion processing are performed on the received data to generate a full-spectrum image of the evidence.

[0049] As a preferred embodiment of the present invention, the activation method of the visible and near-infrared light source in the hyperspectral image acquisition unit is as follows:

[0050] The hyperspectral image acquisition unit determines the target wavelength range, light intensity requirements, color temperature parameters, and core operating parameters according to the preset acquisition task. Based on the requirements of hyperspectral analysis, the target wavelength range of 400nm-760nm is used to capture the color and basic material characteristics of physical evidence, and the target wavelength range of 760nm-1000nm is used to detect the near-infrared absorption characteristics of material molecules.

[0051] The communication control submodule of the hyperspectral image acquisition unit encapsulates the core operating parameters into instruction frames according to a preset protocol format, converts the TTL signal into an RS485 differential signal based on a level conversion circuit, and sends the instruction frames to the communication receiving end of the multispectral light source control module through a shielded twisted pair cable to achieve data transmission. Based on the instruction frames received by the communication receiving end, it performs identification verification, parses the parameter area data, and extracts the core operating parameters. Based on the core operating parameters, it sends a control signal to the driving circuit of the visible and near-infrared light source to start the light source.

[0052] In this embodiment, the hyperspectral image acquisition unit automatically determines three types of core operating parameters based on the specific task of ink trace detection:

[0053] Target wavelength range: 400nm-760nm visible light band, used to capture the color difference between blue-black ink and paper (such as the blue-black tone of the original ink and the grayish characteristics of the altered ink), as well as the basic material reflective properties of paper fibers; 760nm-1000nm near-infrared band, focusing on detecting functional groups in ink molecules (such as the near-infrared absorption peak of ferricyanide), to distinguish the compositional differences between different batches of ink.

[0054] Light intensity requirement: For the diffuse reflection characteristics of A4 documents (210mm×297mm), the preset illuminance is 800 lux, ensuring that the difference in reflected light intensity between the ink-penetrated area and the non-penetrated area is ≥30%, which meets the spectral resolution requirement.

[0055] Color temperature parameters: Select 5500K neutral white light to avoid color interference caused by color temperature shift (such as warm light may mask the subtle color difference of ink), and ensure that the spectral distribution uniformity deviation in the 400nm-760nm band is ≤5%.

[0056] The system's built-in parameter verification module verifies the set values: it checks whether the wavelength range covers the ink characteristic peaks (520nm-580nm blue-black ink absorption peak, 800nm-900nm paper penetration characteristic area) to ensure no key ranges are missed; it calculates the matching degree between light intensity and document material (paper reflectivity is about 85%), and the reflected light intensity can reach 680 lux under 800 lux illuminance, meeting the sensor signal-to-noise ratio requirement of ≥60dB; the color temperature parameter is tested for matching with the standard color card to ensure that the color reproduction error is ≤2, and finally the above parameter combination is locked.

[0057] Specifically, the method for acquiring spectral data in the hyperspectral image acquisition unit is as follows:

[0058] Based on the inspection requirements, the spectral band range is locked. The hyperspectral image acquisition unit automatically transmits the band parameters to the tunable filter group of the pushbroom spectrophotometer. The filters in the tunable filter group complete the initial positioning according to the wavelength requirements. The motorized zoom lens automatically adjusts the focal length according to the size of the evidence and confirms the image clarity through the laser focusing assistance system.

[0059] The linear array sensor of the pushbroom spectrophotometer is aligned with the narrow strip area of ​​the evidence. The incident light is focused by the lens and enters the spectrometer, where it is decomposed into a continuous spectrum of 400nm-1000nm. Light signals of different wavelengths are focused onto different pixel positions of the linear array sensor. After a single exposure, a line of raw spectral data containing 200 bands is generated.

[0060] The scanning drive mechanism drives the camera to move along a direction perpendicular to the linear array. The movement step is matched with the pixel size, and the spectral data of the evidence is collected line by line. The collected spectral data is stitched together in the scanning order to generate a three-dimensional data structure containing spatial and spectral dimensions.

[0061] In this embodiment, based on the requirements for ink trace inspection (distinguishing the differences in pigment composition and paper penetration of different inks), the system automatically locks the core wavelength range. The main analysis wavelength is 500nm-700nm, covering the characteristic absorption peak of ferric ferrocyanide in blue and black inks, with 520nm-580nm being the key range. The auxiliary analysis wavelength is 760nm-900nm, reflecting the differences in the penetration depth of ink in paper fibers, with a more significant decrease in reflectivity in areas of full penetration.

[0062] The hyperspectral image acquisition unit transmits the aforementioned band parameters to the tunable filter array of the pushbroom spectrophotometer via an internal communication protocol. The 10 high-resolution interference filters in the filter array are initially positioned according to wavelength requirements, ensuring a spectral resolution of 2nm in the 500nm-700nm band and 3nm in the 760nm-900nm band. The motorized zoom lens obtains the document size (210mm × 297mm standard A4 paper) through an image recognition module and automatically adjusts the focal length to 150mm to completely cover the document area. The laser focusing assist system emits a 650nm positioning laser. By monitoring the sharpness of the laser spot reflection on the document surface, the focusing evaluation function value is calculated to reach 0.92 (≥0.8 is acceptable), confirming that the image sharpness meets the requirements.

[0063] The pushbroom-type spectrophotometer's linear CCD sensor is aligned with a narrow strip (0.05mm wide) at the top of the document. When this area is illuminated by a visible or near-infrared light source, the reflected light is focused by the lens into a grating beam splitter, where it is decomposed into a continuous spectrum from 400nm to 1000nm. Light signals of different wavelengths are focused sequentially onto different pixel positions of the linear sensor. A scanning drive mechanism moves the camera at a constant speed along a direction perpendicular to the linear array (the document's length), with a step size set to 0.01mm (matching the pixel size) to ensure a 99% spatial overlap between adjacent scan lines. Each step triggers a line scan, acquiring spectral data from the document surface line by line.

[0064] During the acquisition process, the first data preprocessing unit processes each line of spectral data in real time. It subtracts the dark current data acquired in the dark field by dark current subtraction to eliminate sensor thermal noise. For the three abnormal pixels caused by sensor defects, the 5×5 neighborhood mean is used for replacement. After the repair, the data continuity error is greatly reduced.

[0065] Specifically, the method for fusing preprocessed data in the hyperspectral image acquisition unit is as follows:

[0066] A spectral feature extraction network comprising convolutional layers, pooling layers, and residual connections is constructed. The convolutional layers capture multi-scale spectral information of evidence from local details to global structure. The pooling layers reduce data dimensionality while preserving key features. The residual connections enhance the learning ability of the spectral feature extraction network for complex spectral features. A spatial feature extraction network captures multi-scale features of evidence images from microscopic texture details to macroscopic geometric contours.

[0067] By calculating the mutual information matrix of spectral features and spatial features, relevant feature pairs are identified, and the correlation between spectral features and spatial features is established based on the relevant feature pairs. The spectral feature vector and spatial feature vector are multiplied by the corresponding attention weights through an adaptive weighted fusion algorithm to obtain weighted spectral features. The two types of weighted features are fused into a joint feature vector through feature splicing and residual connection to generate a complete physical evidence image-spectrum integrated dataset.

[0068] In this embodiment, a spectral feature extraction network consisting of three convolutional layers, two pooling layers, and residual connections is constructed based on preprocessed three-dimensional hyperspectral data. The first convolutional layer (3×3×16 kernels) captures local spectral details, such as the intensity changes of ink absorption peaks in the 520nm-580nm band, and outputs a 16-channel feature map. The second convolutional layer (3×3×32 kernels) extracts spectral peak shape features to distinguish between blue-black ink (strong absorption at 550nm) and tampered ink (weakened absorption at 550nm). The third convolutional layer (3×3×64 kernels) integrates global spectral trends to identify paper penetration depth features (abnormal reflectance in tampered areas) in the 760nm-900nm band.

[0069] Pooling and residual optimization are employed, with a 2×2 max-pooling layer added after every two convolutional layers to reduce the feature dimension from 200 bands to 50. Simultaneously, residual connections mitigate gradient vanishing, improving the network's accuracy in recognizing subtle spectral differences (such as a 10% difference in absorption peak intensity) to 92%. A 1024-dimensional spectral feature vector is generated, containing quantitative features of ink composition, concentration distribution, and paper penetration.

[0070] For the spatial texture data of full-spectrum imagery (4000×29700 pixels), a spatial feature extraction network based on the U-Net architecture is used for processing: through four convolutions and pooling, spatial features from micro to macro are extracted, including: micro features, 0.1mm-level texture of ink strokes (disorder of strokes in the tampered area); macro features, the ink distribution area of ​​the entire document (boundary contours between the tampered area and the original area).

[0071] The mutual information matrix (size 1024×512) of 1024-dimensional spectral features and 512-dimensional spatial features was calculated to identify highly correlated feature pairs: 550nm absorption peak intensity (spectral feature) and pen stroke texture direction (spatial feature), with a mutual information value of 0.78, reflecting the "absorption peak intensity-pen stroke regularity" correlation of the original ink; 800nm ​​reflectance (spectral feature) and region edge gradient (spatial feature), with a mutual information value of 0.72, corresponding to the "abnormal reflectance-blurred edge" characteristics of the tampered region; feature pairs with mutual information values ​​<0.3 (such as paper background spectrum and document edge) were marked as low weight to reduce noise interference.

[0072] Specifically, the feature splicing and residual connection are implemented as follows:

[0073] The spectral feature vector magnitude is unified by normalization, and the spatial feature vector value is compressed to a preset range by standardization. The hyperspectral image acquisition unit automatically verifies the consistency of the sample quantity of the two types of features. Based on the verification result, the standardized spectral feature vector and the spatial feature vector are directly concatenated along the channel dimension to generate an initial fused feature vector. The initial fused feature vector is input into the residual module, and joint features are formed by standardizing the mean and variance of the feature vector and compressing the dimension. The joint feature vector is associated with the data acquired by the hyperspectral image acquisition unit to output a complete dataset.

[0074] Specifically, the interferogram in the short-wave infrared hyperspectral acquisition unit is formed as follows:

[0075] The short-wave infrared hyperspectral analyzer performs optical path calibration, resets the interferometer moving mirror to the zero optical path difference position, and the laser collimation system emits an auxiliary positioning laser. By monitoring the position deviation of the reflected light spot, the tilt angle of the fixed mirror is adjusted to complete the convergence of the two beams on the target surface of the MCT detector.

[0076] After spectral acquisition is initiated, the servo motor drives the moving mirror of the interferometer to move along the optical axis. Based on the linear change in the reflected optical path difference between the fixed mirror and the moving mirror as the moving mirror moves, the two coherent beams converge at the MCT detector, and interference occurs due to the difference in optical path difference. The MCT detector converts the interference light intensity signal into an analog electrical signal, amplifies it through a preamplifier, and converts it into a digital signal by an analog-to-digital converter. The digital signal is then processed using dark current subtraction, bad pixel repair, and phase correction techniques to output the interferogram.

[0077] In this embodiment, the short-wave infrared light reflected from the document surface is focused by a lens and enters the interferometer. It is then split into two beams by a beam splitter: 50% of the reflected light is reflected by a fixed mirror, and 50% of the transmitted light is reflected by a moving mirror. The two beams produce a linearly varying optical path difference due to the movement of the moving mirror. When the optical path difference Δ = λ (λ is the wavelength), constructive interference (signal enhancement) occurs; when Δ = λ / 2, destructive interference (signal weakening) occurs. The two coherent beams converge at the target surface of an MCT detector (-80℃ cryogenic cooling, quantum efficiency ≥70%). The detector converts the intensity change of the interference light into an analog electrical signal (voltage range 0-5V), with a response time ≤10μs.

[0078] The preamplifier uses a low-noise instrumentation amplifier circuit (gain 100 times, noise ≤1μV) to amplify the weak electrical signal (10mV-500mV) output by the detector to 1V-5V, ensuring that the signal dynamic range covers 60dB; the 16-bit high-speed AD converter digitizes the amplified signal at a sampling rate of 2MHz, generating one data point every 0.5μs, generating 40,000 raw data points in a single scan, completely recording the changes in interference fringes.

[0079] The original digital signal is processed in real time by the preprocessing unit. The processed signal generates an interferogram, with the horizontal axis representing the optical path difference and the vertical axis representing the corrected light intensity value. The interferogram shows a strong peak at the center position with zero optical path difference, which decays sinusoidally to both sides, forming symmetrically distributed interference fringes. The interferogram of the tampered ink area shows that the peak intensity at the optical path difference position at 1450nm is 25% lower than that of the original ink area, and the peak shape at the position corresponding to 2100nm shows obvious distortion.

[0080] Specifically, after the short-wave infrared hyperspectral acquisition unit is started, it performs self-checks on the Fourier transform spectral component, MCT detector, optical adjustment mechanism, and data transmission link. Any abnormalities are displayed in real time through the terminal interaction module. The ultra-low temperature MCT detection module starts a precision temperature control program. Based on proportional-integral-derivative closed-loop control technology, the semiconductor cooler starts a graded cooling mode, which rapidly reduces the temperature of the MCT detector through high-power cooling. When the temperature approaches the target value, it switches to a low-power fine adjustment mode, which works in conjunction with the temperature sensor for real-time monitoring and feedback.

[0081] Specifically, the reflectance conversion in the short-wave infrared hyperspectral acquisition unit is implemented as follows:

[0082] The sample is placed on the stage, and the short-wave infrared hyperspectral analyzer scans it, causing the moving mirror of the interferometer to move and generate an optical path difference. The MCT detector receives the interference signal of the light reflected from the sample and generates the original light intensity spectrum based on the fast Fourier transform. During the optical path calibration stage, the light intensity data reflected by the standard reference plate is collected as the reference plate reference spectrum.

[0083] The original light intensity spectrum is subjected to dark current subtraction processing. Based on the reflectance of the reference plate's baseline spectrum and the processed original light intensity spectrum, a wavelength-specific calibration coefficient is calculated. The calculated calibration coefficient is then smoothed. Based on the calibration coefficient, the relative light intensity is converted to absolute reflectance by performing reflectance conversion on the channels of the original light intensity spectrum.

[0084] In this embodiment, the document to be examined (containing the original blue-black ink and the suspected tampered gray ink area) is fixed on a constant-temperature stage (25℃±0.5℃). The short-wave infrared hyperspectral analyzer starts scanning after receiving the acquisition command. The interferometer's moving mirror moves at a constant speed along the optical axis under the drive of a servo motor, generating a continuous optical path difference of 0-5000nm, covering the 1000nm-2500nm analysis band. The ultra-low temperature MCT detector (cooled to -80℃±1℃) receives the interference signal of the light reflected from the document, converts the light intensity change into an analog electrical signal (voltage range 0-3V), amplifies it by a preamplifier (gain 1000 times), and then converts it into a digital signal by a 16-bit A / D converter (sampling rate 1MHz).

[0085] A Fast Fourier Transform (FFT) is performed on the interference signal to convert the optical path difference domain data into wavelength domain data, generating the original intensity spectrum. Before sample scanning, the optical path calibration module drives the built-in gold-plated standard reference plate into the optical path. A halogen tungsten lamp light source illuminates the reference plate, and the interference signal of its reflected light is collected. After FFT conversion, the original intensity spectrum of the reference plate is generated. The acquisition is repeated 32 times and averaged to reduce noise interference, making the signal-to-noise ratio of the reference spectrum ≥1000:1.

[0086] Specifically, the full-spectrum image acquisition unit supports pseudo-color synthesis, and the synthesis method is as follows:

[0087] Three bands were randomly selected within the 200nm-2500nm range. The reflectance values ​​of each band were linearly stretched to a grayscale value range of 0-255 to perform radiometric normalization on the selected bands. Median filtering was used to eliminate high-frequency noise and preserve edge features. The selected band images were spatially registered using a sub-pixel level registration algorithm.

[0088] Based on the band matching algorithm, the preprocessed selected band data is mapped to the red, green, and blue channels respectively. An adaptive white balance algorithm is used to monitor the spectral distribution characteristics of each pixel in the image in real time, and automatically adjust the color temperature and hue parameters based on the detection results. A combination of piecewise linear stretching and histogram equalization is used to remap the grayscale range of each channel to suppress color overflow. A local color enhancement module analyzes the texture and spectral characteristics of the trace area to enhance color contrast and edge sharpness. Based on the identification of the trace target and background area, the hue and saturation of the target area are optimized using a color mapping table to generate a pseudo-color image.

[0089] In this embodiment, a combination of piecewise linear stretching and histogram equalization is used to expand the grayscale range and enhance weak features. The stretching slope for the low grayscale area (tampered ink) of the blue channel (ultraviolet) is set to 2.0, and the high grayscale area (background) is set to 0.5, making the blue of the tampered area more prominent. The red channel (near infrared) is equalized to expand the low grayscale range (0-70) of the original ink to 0-100, improving the visual recognition of penetration differences (contrast improvement of 40%).

[0090] Local color enhancement is performed on the ink edges and tampering boundaries. The ink edges are detected by the Sobel operator (gradient value ≥ 0.6), and the RGB values ​​of the edge areas are sharpened. The saturation of the tampered areas (UV reflectance ≥ 0.5 and near-infrared reflectance ≥ 0.6) is increased by 20% to make the colors more vivid and distinguish them from the background by ≥ 80%.

[0091] Based on the trace recognition results, the color mapping table is optimized. In the pseudo-color, it presents a mixed orange-yellow color of "red + green" (weak near-infrared reflection → light red, strong visible light absorption → light green); due to strong ultraviolet reflection (deep blue), strong visible light reflection (deep green), and strong near-infrared reflection (deep red), it presents a mixed cyan-purple color of "blue + green + red", which forms a significant color difference with the original area.

[0092] The generated pseudo-color image has the following characteristics: 20 million pixels (6000×4000), clearly presenting ink stroke details down to 0.1mm level; the tampered area (2cm×3cm) is highlighted in cyan-purple, with complete edge contours (overlap with the original area boundary ≥98%); the two types of ink can be intuitively distinguished through pseudo-color comparison, and combined with spectral data verification, the accuracy of tampered area identification reaches 100%, providing visual basis for subsequent qualitative analysis.

[0093] Specifically, the steps for the ultraviolet-enhanced lens and the anti-reflective coating to work together in the full-spectrum image acquisition unit are as follows:

[0094] Full-band light carrying physical evidence information enters the imaging component, and the ultraviolet enhancement lens based on the quartz glass substrate controls the attenuation of ultraviolet light; based on the optical interference structure, the outer film system of the anti-reflection coating suppresses the reflection of light in each band.

[0095] Based on low-absorption special optical materials, the ultraviolet-enhanced lens reduces ultraviolet light propagation loss at the substrate level; the anti-reflective coating achieves deep suppression of interface reflection through a multi-layer gradient refractive index ultraviolet-specific film system design; the aspherical geometry of the ultraviolet-enhanced lens and the refractive index modulation of the anti-reflective coating form an optical compensation mechanism, which corrects aberrations and focusing deviations in the ultraviolet band based on matching the refractive characteristics of ultraviolet and visible light, achieving sub-pixel-level spatial alignment between ultraviolet imaging and visible light imaging.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A unified intelligent examination system based on ultrawide spectrum and physical evidence images, characterized in that, include: A hyperspectral image acquisition unit is used to acquire continuous spectral images of physical evidence in the 400nm-1000nm band. It sends a working band request to the multispectral light source control module via a communication interface, triggering the activation of the visible and near-infrared light source. The multispectral light source control module presets the initial illuminance based on the material of the physical evidence. The pushbroom-type spectral camera uses a linear pushbroom mode, with an internal scanning mechanism driving the spectral sensor to move along the surface of the physical evidence to complete spectral data acquisition. The acquired data is preprocessed based on dark current correction, spectral calibration, and bad pixel repair techniques. The preprocessed data is then fused by an image fusion processing module to output a unified physical evidence image dataset. The method for fusing the preprocessed data in the hyperspectral image acquisition unit is as follows: A spectral feature extraction network comprising convolutional layers, pooling layers, and residual connections is constructed. The convolutional layers capture multi-scale spectral information of evidence from local details to global structure. The pooling layers reduce data dimensionality while preserving key features. The residual connections enhance the learning ability of the spectral feature extraction network for complex spectral features. A spatial feature extraction network captures multi-scale features of evidence images from microscopic texture details to macroscopic geometric contours. By calculating the mutual information matrix of spectral features and spatial features, relevant feature pairs are identified, and the correlation between spectral features and spatial features is established based on the relevant feature pairs. The spectral feature vector and spatial feature vector are multiplied by the corresponding attention weights through an adaptive weighted fusion algorithm to obtain weighted spectral features. The two types of weighted features are fused into a joint feature vector through feature concatenation and residual connection to generate a complete physical evidence image-spectrum integrated dataset. The feature splicing and residual connection are implemented as follows: The spectral feature vectors are normalized to unify their magnitudes, and the spatial feature vectors are compressed to a preset range through standardization. The hyperspectral image acquisition unit automatically verifies the consistency of the sample size of the two types of features. Based on the verification result, the standardized spectral feature vectors and spatial feature vectors are directly concatenated along the channel dimension to generate an initial fused feature vector. The initial fused feature vector is input into the residual module, and joint features are formed by standardizing the mean and variance of the feature vectors and compressing their dimensions. The joint feature vector is then correlated with the data acquired by the hyperspectral image acquisition unit to output a complete dataset. The short-wave infrared hyperspectral acquisition unit is used to acquire hyperspectral data of physical evidence in the 1000nm-2500nm band; the short-wave infrared hyperspectral analyzer sends the short-wave infrared band request to the multi-band light source system through the communication protocol, triggers the halogen tungsten lamp light source to start, and the optical path calibration module drives the built-in standard reference board into the optical path to acquire the reference spectrum. The moving mirror of the interferometer moves under the drive of a motor, generating a change in optical path difference, causing short-wave infrared light of different wavelengths to interfere and form an interferogram; the interferogram is corrected, wavelength calibrated and reflectivity converted, and the short-wave infrared spectrum of the physical evidence is output; A full-spectrum image acquisition unit is used to acquire broadband images of the evidence in the 20nm-2500nm wavelength range; a full-spectrum image analyzer sends a full-spectrum wavelength request to the multi-band light source system, triggering the multi-spectral LED array light source to start and configuring the initial color temperature and initial illuminance; an optical path calibration module drives a standard color chart into the field of view to acquire calibration images; the ultraviolet enhancement lens and anti-reflection coating in the broadband imaging component work together to receive the full-spectrum light reflected by the evidence through a CMOS sensor, capturing color information and differences in infrared reflection of the material; noise removal, color restoration, and band fusion processing are performed on the received data to generate a full-spectrum image of the evidence.

2. The system according to claim 1, characterized in that, The method for activating the visible and near-infrared light source in the hyperspectral image acquisition unit is as follows: The hyperspectral image acquisition unit determines the target wavelength range, light intensity requirements, color temperature parameters, and core operating parameters according to the preset acquisition task. Based on the requirements of hyperspectral analysis, the target wavelength range of 400nm-760nm is used to capture the color and basic material characteristics of physical evidence, and the target wavelength range of 760nm-1000nm is used to detect the near-infrared absorption characteristics of material molecules. The communication control submodule of the hyperspectral image acquisition unit encapsulates the core operating parameters into instruction frames according to a preset protocol format, converts the TTL signal into an RS485 differential signal based on a level conversion circuit, and sends the instruction frames to the communication receiving end of the multispectral light source control module through a shielded twisted pair cable to achieve data transmission. Based on the instruction frames received by the communication receiving end, it performs identification verification, parses the parameter area data, and extracts the core operating parameters. Based on the core operating parameters, it sends a control signal to the driving circuit of the visible and near-infrared light source to start the light source.

3. The system according to claim 1, characterized in that, The method for acquiring spectral data in the hyperspectral image acquisition unit is as follows: Based on the inspection requirements, the spectral band range is locked. The hyperspectral image acquisition unit automatically transmits the band parameters to the tunable filter group of the pushbroom spectrophotometer. The filters in the tunable filter group complete the initial positioning according to the wavelength requirements. The motorized zoom lens automatically adjusts the focal length according to the size of the evidence and confirms the image clarity through the laser focusing assistance system. The linear array sensor of the pushbroom spectrophotometer is aligned with the narrow strip area of ​​the evidence. The incident light is focused by the lens and enters the spectrometer, where it is decomposed into a continuous spectrum of 400nm-1000nm. Light signals of different wavelengths are focused onto different pixel positions of the linear array sensor. After a single exposure, a line of raw spectral data containing 200 bands is generated. The scanning drive mechanism moves the camera along a direction perpendicular to the linear array, with the movement step size matching the pixel size. It collects spectral data of the evidence line by line, and the collected spectral data is stitched together in the scanning order to generate a three-dimensional data structure containing spatial and spectral dimensions.

4. The system according to claim 1, characterized in that, The interferogram in the short-wave infrared hyperspectral acquisition unit is formed as follows: The short-wave infrared hyperspectral analyzer performs optical path calibration, resets the interferometer moving mirror to the zero optical path difference position, and the laser collimation system emits an auxiliary positioning laser. By monitoring the position deviation of the reflected light spot, the tilt angle of the fixed mirror is adjusted to complete the convergence of the two beams on the target surface of the MCT detector. After spectral acquisition is initiated, the servo motor drives the moving mirror of the interferometer to move along the optical axis. Based on the linear change in the reflected optical path difference between the fixed mirror and the moving mirror as the moving mirror moves, the two coherent beams converge at the MCT detector, and interference occurs due to the difference in optical path difference. The MCT detector converts the interference light intensity signal into an analog electrical signal, amplifies it through a preamplifier, and converts it into a digital signal by an analog-to-digital converter. The digital signal is then processed using dark current subtraction, bad pixel repair, and phase correction techniques to output the interferogram.

5. The system according to claim 1, characterized in that, After the short-wave infrared hyperspectral acquisition unit is started, it performs self-checks on the Fourier transform spectral component, MCT detector, optical adjustment mechanism, and data transmission link. Any abnormalities are displayed in real time through the terminal interaction module. The ultra-low temperature MCT detection module starts a precision temperature control program. Based on proportional-integral-derivative closed-loop control technology, the semiconductor cooler starts a graded cooling mode, which rapidly reduces the temperature of the MCT detector through high-power cooling. When the temperature approaches the target value, it switches to a low-power fine adjustment mode, which works in conjunction with the temperature sensor for real-time monitoring and feedback.

6. The system according to claim 1, characterized in that, The reflectivity conversion in the shortwave infrared hyperspectral acquisition unit is implemented as follows: The sample is placed on the stage, and the short-wave infrared hyperspectral analyzer scans it, driving the moving mirror of the interferometer to move and generate an optical path difference. The MCT detector receives the interference signal of the light reflected from the sample and generates the original light intensity spectrum based on the fast Fourier transform. During the optical path calibration phase, the light intensity data reflected from the standard reference plate is collected as the reference plate's baseline spectrum. The original light intensity spectrum is subjected to dark current subtraction processing. Based on the reflectance of the reference plate's baseline spectrum and the processed original light intensity spectrum, a wavelength-specific calibration coefficient is calculated. The calculated calibration coefficient is then smoothed. Based on the calibration coefficient, the relative light intensity is converted to absolute reflectance by performing reflectance conversion on the channels of the original light intensity spectrum.

7. The system according to claim 1, characterized in that, The full-spectrum image acquisition unit supports pseudo-color synthesis, and the synthesis method is as follows: Three bands were randomly selected within the 200nm-2500nm range. The reflectance values ​​of each band were linearly stretched to a grayscale value range of 0-255 to perform radiometric normalization on the selected bands. Median filtering was used to eliminate high-frequency noise and preserve edge features. The selected band images were spatially registered using a sub-pixel level registration algorithm. Based on the band matching algorithm, the preprocessed selected band data is mapped to the red, green, and blue channels respectively; an adaptive white balance algorithm is used to monitor the spectral distribution characteristics of each pixel in the image in real time, and the color temperature and hue parameters are automatically adjusted based on the detection results; based on the combination of piecewise linear stretching and histogram equalization, the gray value range of each channel is remapped to suppress color overflow. The local color enhancement module analyzes the texture and spectral characteristics of the trace area to enhance color contrast and edge sharpness. Based on the identification of the trace target and background area, the color map table is used to optimize the hue and saturation of the target area to generate a pseudo-color image.

8. The system according to claim 1, characterized in that, The steps for the UV-enhanced lens and anti-reflective coating to work together in the full-spectrum image acquisition unit are as follows: Full-band light carrying physical evidence information enters the imaging component, and the ultraviolet enhancement lens based on the quartz glass substrate controls the attenuation of ultraviolet light; based on the optical interference structure, the outer film system of the anti-reflection coating suppresses the reflection of light in each band. Based on low-absorption special optical materials, the ultraviolet-enhanced lens reduces ultraviolet light propagation loss at the substrate level; the anti-reflective coating achieves deep suppression of interface reflection through a multi-layer gradient refractive index ultraviolet-specific film system design; the aspherical geometry of the ultraviolet-enhanced lens and the refractive index modulation of the anti-reflective coating form an optical compensation mechanism, which corrects aberrations and focusing deviations in the ultraviolet band based on matching the refractive characteristics of ultraviolet and visible light, achieving sub-pixel-level spatial alignment between ultraviolet imaging and visible light imaging.

Images