Linearly graded filter smart variable bandwidth hyperspectral imaging device and method

Abstract

The application discloses a linearly graded filter intelligent variable bandwidth hyperspectral imaging device and method. The device comprises, in sequence along the light path incident direction, an imaging objective, a graded filter assembly, a telecentric imaging lens group and a camera; during operation, a one-dimensional linear displacement table drives the graded filter assembly to perform linear scanning motion in a direction perpendicular to the optical axis, so that light in the imaging field of view is transmitted through different center wavelength band-pass filter regions on the graded filter in time. After filtering, monochromatic or narrow-band light is projected onto the photosensitive target surface of the camera through the telecentric imaging lens group, and multiple frames of spectral images of different wavebands are continuously collected. After processing and reconstruction, the collected sequence images form a hyperspectral three-dimensional data cube containing spatial information and spectral information. The application introduces a Grad-CAM driven hardware-algorithm collaborative control mechanism, thereby achieving high-precision, low-cost and rapid spectral data collection while maintaining a high light flux.

Description

Technical Field

[0001] This invention belongs to the field of spectral imaging technology, and specifically relates to a hardware-algorithm collaborative reconstruction method for intelligent variable bandwidth hyperspectral imaging devices using linear variable filters (LVF). Background Technology

[0002] Hyperspectral imaging, as a detection method integrating image and spectral information, can acquire the two-dimensional spatial morphology and one-dimensional spectral characteristics of target objects in a continuous narrow band, thereby enabling in-depth analysis and accurate identification of the target material composition. With the development of sensor technology and optoelectronic materials, hyperspectral imaging has expanded from laboratory research to various practical applications such as agricultural remote sensing, industrial automation detection, food safety sorting, and biomedical diagnostics. However, achieving high light throughput, high acquisition speed, and low system cost while maintaining high spatial resolution remains a core requirement in the field of hyperspectral detection technology.

[0003] In agriculture and food safety, hyperspectral imaging technology can capture subtle spectral changes in crops, providing early warnings of pests and diseases before they are visible to the naked eye. By analyzing reflectance in specific wavelengths, the system can accurately assess chlorophyll content, water content, and nitrogen fertilizer levels in vegetation, providing data support for precision fertilization and irrigation. In food processing lines, this technology enables non-contact detection of the internal quality of agricultural products, such as sugar content grading of fruits, mold screening of grains, and component analysis of meat products, significantly improving the efficiency and accuracy of detection.

[0004] In the fields of industrial inspection and resource recycling, hyperspectral imaging, with its "image and spectrum integration" advantage, has become a core tool for identifying complex components. In lithium battery manufacturing, this technology can be used to detect the uniformity of electrode coatings and identify surface defects. In waste resource recycling, traditional machine vision struggles to distinguish between engineering plastics (such as PE, PP, and PET) that have similar colors but vastly different chemical compositions, while hyperspectral imaging can achieve highly pure automatic sorting based on its unique "spectral fingerprint." Furthermore, this technology also demonstrates irreplaceable application potential in textile composition identification and cultural relic restoration.

[0005] In the biomedical and pharmaceutical fields, hyperspectral imaging offers new avenues for non-invasive diagnosis. Because different tissues and lesion areas exhibit varying absorption and scattering characteristics to specific wavelengths of light, this technology can be used to assist in early screening for skin cancer, monitoring blood oxygen saturation during surgery, and precisely defining tumor margins. Particularly in intensive care unit (ICU) settings, critically ill patients with sepsis and septic shock often experience involuntary tremors or mechanical vibrations from ventilators. Traditional hyperspectral systems (such as LCTF or push-broom systems) typically require over 10 seconds for full-spectrum acquisition, are highly susceptible to motion artifacts, and suffer from distortion of microcirculation parameters, failing to meet the timeliness requirements of rapid bedside triage. In pharmaceutical manufacturing, hyperspectral scanning of tablets or powders allows for real-time monitoring of the uniformity of active ingredient distribution and impurity content, ensuring drug quality meets standards. However, these fields place extremely high demands on the real-time performance and light throughput of imaging devices, and existing imaging equipment often struggles to achieve a balance between high performance and low cost.

[0006] Currently, mainstream hyperspectral imaging solutions on the market all have significant technical limitations in practical applications. Time-division imaging technology, represented by liquid crystal tunable filters (LCTFs), can quickly switch bands through electronic control, but its principle of polarization interference results in extremely low peak transmittance and light energy utilization of less than 20%. This requires the system to significantly extend the exposure time in low-light environments, greatly limiting the detection capability of dynamic targets. At the same time, the manufacturing process of such core components is complex and the purchase cost is expensive, making it difficult to widely adopt them on industrial production lines.

[0007] Another common pushbroom spectral imaging technique uses slits and dispersive elements (such as gratings or prisms) for beam splitting. Although it offers high spectral resolution, the physical slits significantly reduce the light flux entering the detector. Furthermore, factors such as grating diffraction efficiency and lens surface reflectivity further reduce the light flux. To obtain a sufficiently strong signal, the system often requires an extremely high-power light source or an expensive cooled camera, resulting in a bulky and costly imaging device. It also demands extremely precise alignment of the optical path, has poor shock resistance, and is difficult to adapt to complex industrial environments.

[0008] In recent years, linear variable filters (LVFs) have gradually become a research hotspot in the field of spectral imaging due to their high transmittance (nearly 90%) and excellent filtering characteristics. In existing technologies, LVFs are often directly integrated or bonded to the surface of the image sensor's photosensitive chip. While this design reduces size, it suffers from several insurmountable drawbacks: First, the physical gap between the photosensitive chip and the filter leads to severe spectral aliasing and crosstalk between pixels, reducing the purity of the spectral data. Second, this tightly coupled design makes the imaging system inflexible, preventing users from changing objectives or adjusting magnification for different detection targets. Furthermore, limited by chip size, the spectral adjustment range and sampling step size of such devices are fixed after leaving the factory, making it difficult to meet diverse research and industrial needs. Moreover, existing LVF scanning architectures generally lack a reverse guidance mechanism for hardware acquisition strategies, failing to dynamically optimize the physical scan path based on the core spectral characteristics of a specific diagnostic task. This results in a technical contradiction in rapid clinical triage scenarios: "excessive time spent on full-spectrum acquisition" versus "loss of critical information due to narrow-band truncation."

[0009] Therefore, there is an urgent need to develop a scanning high-throughput hyperspectral imaging system that is fast in imaging, has high spectral fidelity, flexible in structure, and low in cost. Summary of the Invention

[0010] To address the technical problems of existing hyperspectral imaging devices, such as the extremely low transmittance and high cost of liquid crystal tunable filters (LCTFs), the limited light throughput of traditional pushbroom systems, and the severe spectral aliasing and lack of flexibility of bonded LVFs, this invention provides a smart variable bandwidth hyperspectral imaging device and method with a linear graded filter. This invention aims to achieve high-precision, low-cost, and rapid spectral data acquisition while maintaining extremely high light throughput by using an independent linear graded filter scanning component in conjunction with a telecentric optical structure and introducing an interpretable deep learning (Grad-CAM) driven hardware-algorithm collaborative control mechanism.

[0011] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A smart variable bandwidth hyperspectral imaging device with a linear gradient filter includes: An imaging objective, a graduated filter assembly, a telecentric imaging lens group, and a camera are arranged sequentially along the incident direction of the light path. The gradient filter assembly is mounted on a one-dimensional linear displacement stage, and the displacement stage drives the gradient filter assembly to perform linear scanning motion in a direction perpendicular to the optical axis of the system. The gradient filter assembly has a center wavelength transmittance characteristic that varies linearly along its surface spatial position, and is used to split the incident beam. The telecentric imaging lens group is used to project the split beam onto the photosensitive target surface of the camera to ensure that the main ray is incident parallel to the optical axis and to ensure the consistency of spectral characteristics at each position of the filter. The camera continuously acquires images at a high frame rate during the scanning process, and each frame image corresponds to a scanning position of the gradient filter assembly and the corresponding spectral band. The one-dimensional linear displacement stage is equipped with a position feedback sensor to provide real-time feedback on the current position of the gradient filter assembly, thereby establishing a mapping relationship between the scanning position and the center wavelength. The spectral calibration process for hyperspectral data includes: (i) Remove the imaging objective lens and illuminate the device using a wide-field standard characteristic spectrum light source; (ii) Extract the pixel width coordinate w of the feature spectral line from the image corresponding to the position x of each one-dimensional linear displacement stage, and combine it with the known standard wavelength λ0 to establish the wavelength function relationship λ(x,w) = a·x + b·w +c using the least squares method; (iii) The data reconstruction process includes: calculating the actual center wavelength corresponding to each pixel of each frame image according to the wavelength function λ(x,w); performing spatial alignment and stitching of the image sequence based on the calculation results; setting a fixed spectral coordinate grid, and reconstructing the spectral information of all pixels to the grid through an interpolation algorithm to generate a hyperspectral three-dimensional data cube; The device also includes an intelligent scanning control unit: The intelligent scanning control unit is electrically connected to the one-dimensional linear displacement stage and the camera, and is configured to: acquire a pre-trained attention weight model for a specific classification task; extract the spectral feature attention distribution of the target object based on the model, and identify the core continuous band interval with high contribution and the background band interval with low contribution; The intelligent scanning control unit maps the core band interval and background band interval to the physical start and end coordinates of a one-dimensional linear displacement stage. It performs truncated scanning only within the effective stroke formed by the physical start and end coordinates. Based on the truncated scan frame sequence, it reconstructs a truncated hyperspectral data cube that removes redundant background bands and inputs it into the classification decision model for retraining to complete classification or prediction, thereby improving the acquisition and recognition speed and suppressing motion artifacts.

[0012] The device also includes a bandwidth adjustment mechanism for continuously fine-tuning the synthesized spectral bandwidth by adjusting the relative position or transmission characteristics of the filters inside the gradient filter assembly. The gradient filter assembly includes a linearly gradient long-pass filter and a linearly gradient short-pass filter stacked in parallel along the optical axis. The two filters change in the same direction and have a peak transmittance of not less than 80%. They form an adjustable bandwidth bandpass window by adjusting their relative positions. The bandwidth adjustment mechanism is used to adjust the relative positions of the two filters to achieve continuous fine-tuning of the bandwidth in the range of 5 nm to 30 nm.

[0013] The gradient filter assembly is a single linear gradient bandpass filter with a transmission center wavelength that changes linearly and continuously along one geometric axis of the filter. The wavelength gradient range is 400 nm to 1000 nm, the full width at half maximum (FWHM) is 6 to 20 nm, and the peak transmittance is not less than 80%.

[0014] The telecentric imaging lens group is an object-oriented telecentric lens or a double telecentric lens, whose object plane position matches the center position of the graduated filter assembly to ensure that the incident principal ray is parallel to the optical axis, and the optical system error of the lens group is controlled within λ / 10.

[0015] The camera is a high frame rate area array detector, which is electrically connected to a one-dimensional linear displacement stage through a synchronization control unit. During the scanning process, the camera triggers acquisition at a preset frame rate to achieve precise synchronization between image acquisition and filter position. The one-dimensional linear displacement stage includes a precision electric slide and a position feedback sensor. The sensor is used to provide real-time feedback on the current position of the gradient filter assembly and to support wavelength positioning and data reconstruction.

[0016] The bandwidth adjustment mechanism achieves fine-tuning of the synthesized spectral bandwidth by adjusting the relative position or transmission characteristics of each filter inside the gradient filter assembly. The adjustment range is 5 nm to 30 nm, and the adjustment step is ≤10 μm.

[0017] The process of establishing the wavelength function λ(x,w) also includes: performing multi-frame averaging on the calibration image sequence to reduce noise and improve fitting accuracy, with a fitting residual ≤2 nm.

[0018] The device is suitable for applications such as microcirculation monitoring in intensive care units (ICUs) for sepsis and septic shock, tissue blood oxygenation imaging, shortwave infrared identification of medicinal materials, agricultural remote sensing, food safety testing, or industrial sorting.

[0019] A method for acquiring hyperspectral data using the aforementioned device includes the following steps: S1: The light from the target scene is introduced into the optical path through the imaging objective lens; S2: Control the one-dimensional linear displacement stage to drive the gradient filter assembly to perform linear scanning in the direction perpendicular to the optical axis. The speed is dynamically matched according to the 60 fps frame rate and 2 nm step size to achieve full spectrum acquisition in 5 seconds or acquisition of a custom truncation spectral range. S3: During the scanning process, the camera synchronously acquires a series of two-dimensional narrowband images through the filter at a preset frame rate ≥60 fps. Each frame corresponds to the scanning position x and the center wavelength λ. S4: Record the scanning position x corresponding to each frame of the image, and determine the center wavelength corresponding to each frame of the image based on the linear dispersion characteristics of the gradient filter and the wavelength function λ(x,w). S5: Spatial alignment and spectral interpolation reconstruction are performed on the acquired image sequence to generate a hyperspectral three-dimensional data cube containing spatial and spectral information. The non-uniformly distributed spectral information is mapped to a standard grid with a step size of 2 nm using a cubic spline interpolation algorithm.

[0020] Step S2 further includes an adaptive truncation scan control process based on an AI attention mechanism: S2-1: Intelligent division of core band intervals: Using a pre-trained classification decision model combined with an interpretability algorithm, the attention weight distribution of the model in the full spectrum dimension is extracted; based on the set weight threshold, one or more core continuous spectral band intervals that contribute the most to the decision of the current classification task are automatically determined. S2-2: Control parameter reverse physical mapping: Using the wavelength function λ(x,w)= a·x + b·w + c described in the spectral calibration process of hyperspectral data, the wavelength boundaries of the core continuous spectral band interval are directly reverse mapped to the physical start and end position coordinates of a one-dimensional linear displacement stage. S2-3: Truncation Scan Execution and Data Reconstruction: The one-dimensional linear displacement stage is controlled to directly position itself to the physical starting position coordinates, and truncated scanning is performed only within the effective travel range formed by the physical starting position coordinates and the ending position coordinates; the camera is triggered by the position feedback of the one-dimensional linear displacement stage during the truncated scanning and synchronously acquires the corresponding frame sequence; acquisition stops after scanning to the ending position coordinates, and scanning of non-core band intervals is completely skipped; based on the truncated scan frame sequence, a truncated hyperspectral data cube with redundant background bands removed is reconstructed and input into the classification decision model to complete classification or prediction.

[0021] Compared with the prior art, the beneficial effects of the present invention are reflected in: High luminous flux and high transmittance: Using a linear gradient filter as the beam splitter, its peak transmittance can reach 80%-90%, which is far higher than LCTF technology. This greatly shortens the exposure time of a single frame image and improves the system's imaging capability in low-light environments.

[0022] High spectral accuracy and no aliasing: By placing the filter between the objective lens and the telecentric imaging lens group, and combining it with the telecentric optical path design, the beam is ensured to be perpendicularly incident, eliminating pixel crosstalk and spectral aliasing problems caused by traditional bonding designs. This expands the combination matching of filters of different sizes with cameras, and the spectral resolution can be further improved by using longer filters.

[0023] Accurate data reconstruction: With the help of precision displacement stage feedback and calibration algorithm, wavelength shift during the scanning process can be eliminated, and the generated reconstructed data has extremely high spatial and spectral fidelity.

[0024] Hardware-algorithm collaboration combats motion artifacts, with a flexible structure and low cost: The device adopts a modular design, eliminating the need for customized chip bonding processes. Different magnification objectives or cameras can be replaced as needed, and it avoids expensive photoacoustic or liquid crystal tuning components, significantly reducing hardware costs.

[0025] By introducing Grad-CAM interpretable AI-guided physical stroke cutoff of the displacement stage, the single acquisition time is reduced from 5 seconds to 3 seconds while maintaining clinical classification accuracy of over 91%, completely eliminating motion artifacts caused by tremors in critically ill ICU patients and achieving rapid bedside triage with "instant scanning and diagnosis". Attached Figure Description

[0026] Figure 1 : Schematic diagram of the optical path structure of the device of the present invention; In the figure, there are: 1. Imaging objective lens; 2. One-dimensional linear displacement stage; 3. Linearly graded long-pass filter; 4. Linearly graded short-pass filter; 5. Telecentric imaging lens group; 6. Camera; 7. Bandwidth adjustment mechanism.

[0027] Figure 2 : Flowchart of spectral calibration and data reconstruction.

[0028] Figure 3 Visible-near-infrared tissue oxygenation imaging: hyperspectral imaging of the palm.

[0029] Figure 4 Visible-near-infrared tissue oxygenation imaging image.

[0030] Figure 5 Visible-near-infrared tissue hemoglobin index imaging.

[0031] Figure 6 : A schematic diagram of the attention weight distribution and core continuous spectral band intervals of tissue features extracted using the Grad-CAM algorithm.

[0032] Figure 7 The classification confusion matrix is ​​obtained by retraining after reconstructing the data based on the core spectral bands.

[0033] Figure 8 The classification ROC curve obtained by retraining based on the data reconstructed from the core spectral bands.

[0034] Figure 9 Photograph of a sample of Panax notoginseng.

[0035] Figure 10 Short-wave infrared reflectance spectrum image.

[0036] Figure 11 Short-wave infrared reflectance spectrum curve.

[0037] Figure 12 Photo of the prototype of the hyperspectral analyzer for rapid identification of medicinal materials in operation. Detailed Implementation

[0038] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. Example 1: Visible-Near Infrared Rapid Monitoring Device for Microcirculation in Critical Illnesses

[0039] This embodiment provides a device specifically designed for collecting microcirculation data and analyzing pathological parameters of the palms of ICU patients with sepsis and septic shock. Its core design employs a linearly graded beam splitter system covering the 400-1000nm wavelength range, enabling continuous and efficient spectral imaging from the visible to near-infrared spectrum. Addressing the challenge of motion artifacts caused by involuntary tremors in critically ill patients, this embodiment innovatively introduces an interpretable deep learning-driven hardware-algorithm collaborative truncation scanning strategy.

[0040] like Figure 1 As shown, this embodiment uses the following configuration: (1) Imaging objective lens 1: 25 mm focal length C-mount target lens (MVL-KF2524M-25MP); (2) Graded-pass filter assembly: It is composed of a linearly graded long-pass filter 3 and a linearly graded short-pass filter 4 stacked together with matched parameters. The two achieve linear tuning of the spectral edge in the range of 400-1000 nm, and the linear gradient rate is 20 nm / mm. This assembly forms a movable bandpass window by adjusting the overlapping position, and thanks to the high transmittance of a single piece >80% and the deep cutoff characteristics of >OD5, it effectively ensures the high light throughput and signal-to-noise ratio of the system. The measured FWHM is about 6.63 nm. The bandwidth adjustment mechanism 7 is a manual fine-tuning screw with an adjustment range of 5-30 nm; (3) One-dimensional linear displacement stage 2: driven by a linear motor, with a stroke of 100 mm and a position feedback accuracy of ≤10 μm; (4) Telecentric Imaging Lens Group 5: Object-side telecentric lens, visible-near infrared dual telecentric relay lens (1-inch target surface, 0.71× magnification, 5.3 μm pixel resolution), object plane matched with filter center plane (error ≤ 0.1 mm). (5) Camera: 6: 1-inch CMOS area array detector (ME2P-530-72U3MNIR) with a response band of 400-1000nm, frame rate locked at 60 fps, minimum exposure ≤15 ms, and gain of 8 dB; (6) Synchronous control unit: a collaborative control software based on the Python / PyTorch framework, integrating the displacement table motion command issuance, camera triggering and Grad-CAM reverse mapping algorithm modules.

[0041] like Figure 1 As shown, at the optical front end, the imaging objective 1 focuses the light reflected from the target palm onto the LVF module. The innovation of the beam splitter module lies in its graded filter array, which is not a single filter, but rather a precisely parallel stack of a linearly graded long-pass filter 3 and a linearly graded short-pass filter 4. By finely adjusting the relative starting positions of the two, the overlap area of ​​their transmission curves can be precisely controlled, thereby stably locking the full width at half maximum (FWHM) of the synthesized effective bandpass window at approximately 6.63 nm, forming a movable, extremely narrow spectral screening "window".

[0042] The filter array is integrated onto a one-dimensional linear displacement stage 2. During data acquisition, the stage moves the stacked filter array laterally at a uniform speed, ensuring that its gradient direction aligns with the direction of motion. A telecentric imaging lens group 5 is specifically positioned in front of the camera's photosensitive area as a relay optical system. Its key function is to ensure that light reflected from the tissue and passing through the objective lens is incident on the gradient filter at an approximately perpendicular angle. This dual-telecentric architecture completely eliminates spectral shift and pixel crosstalk caused by large-angle incident light, laying a solid physical data foundation for subsequent high-precision AI feature extraction. This design minimizes center wavelength drift caused by changes in the incident angle, ensuring the accuracy of the spectral data.

[0043] The acquisition action of the area-array CMOS camera 6 is triggered by a high-precision encoder on the displacement stage. During the uniform scanning process of the filter array, the camera rapidly captures a series of instantaneous narrowband images at an extremely high frame rate, each image corresponding to a specific narrow wavelength of light. Due to the overall transmittance of the entire optical path, especially the customized filter array, which is over 80%, the system is able to capture extremely weak spectral fluctuation signals in skin reflection. Full-spectrum (400-1000 nm) scanning completes 300 frame acquisitions within 5 seconds, effectively freezing patient micro-movements.

[0044] like Figure 2 The data acquisition and reconstruction process shown 1. Spectral Calibration: The imaging objective lens was removed, and the illumination device was calibrated using the HG-1 mercury-argon standard calibrator. A sequence of 200 images was acquired, and the pixel width *w* of the characteristic spectral lines was extracted. Combined with the stage position *x*, a wavelength function λ(x,w) = a·x + b·w + c (fitting residual ≤ 2 nm) was established using the least squares method. Here, *a* represents the wavelength gradient rate (nm / mm) of the linearly graded filter, *b* corrects for spatial wavelength differences within the detector's field of view (nm / pixel), and *c* is the initial wavelength intercept of the system. In actual calculations, multiple sets of λ(x,w) data pairs were obtained by identifying characteristic peaks of multiple known standard wavelengths in the light source. Parameters *a*, *b*, and *c* were solved using multiple linear regression to establish a precise mapping model from the "mechanical-pixel" space to the "wavelength" space. Standard grid reconstruction was performed using cubic spline interpolation.

[0045] 2. Data Acquisition: Restore the imaging objective lens, control the stage to scan at a constant speed (matching 60fps and 2nm step size), and the camera simultaneously acquires 300 frames of narrowband images at 60fps. The system gain is set to 8 dB, and the single scan time is strictly limited to ≤5 s; if macroscopic tearing or spatial misalignment is detected, the scan is deemed invalid and a rescan mechanism is triggered.

[0046] 3. Data Reconstruction: Calculate the actual wavelength corresponding to each pixel in each frame of the image based on λ(x,w), perform spatial alignment (column interpolation) on the image sequence, set the spectral grid (400-1000 nm, step size 2 nm), and reconstruct the hyperspectral three-dimensional data cube using cubic spline interpolation.

[0047] 4. Experimental Results: With a data acquisition time of 5 seconds, the spatial resolution of blood oxygen saturation was pixel-level, and the spectral reconstruction error was ≤2nm. The calibrated measured FWHM reached 6.63 nm, verifying the ability of the telecentric optical path to synergistically optimize high resolution and high throughput.

[0048] Subsequent data processing steps, such as Figure 3 , Figure 4 , Figure 5As shown, by reconstructing and calculating the acquired hyperspectral image data of the palm, quantitative physiological parameters such as tissue oxygen saturation (StO2), tissue hemoglobin index (THI), and tissue water index (TWI) can be extracted (the formulas are derived based on the modified Beer-Lambert law), and a high-precision pseudo-color spatial distribution map is output. Clinical statistics show that the sepsis group has significantly reduced StO2, abnormally increased THI and TWI, and an "asynchronous deterioration" phenomenon in individual indicators, confirming the necessity of full-spectrum high-dimensional feature fusion. The reflectance ratio of characteristic bands such as 660nm (hemoglobin-sensitive band) and 940nm (oxyhemoglobin-sensitive band) is also calculated. Based on the bio-optical model, this ratio is used for inversion calculation, ultimately achieving intuitive analysis and imaging of the spatial distribution of subcutaneous blood oxygen saturation and hemoglobin index.

[0049] Compared to traditional liquid crystal tunable filter (LCTF) solutions, the mechanical scanning linear gradient filtering method employed in this embodiment reduces the imaging time of a single hyperspectral data cube by more than 70%. This high-speed acquisition characteristic significantly reduces image registration errors and motion artifacts caused by unavoidable minor subject movements, greatly improving the reliability and practicality of the measurement. Task-driven truncation scan rapid identification system based on AI attention mechanism

[0050] Existing push-broom hyperspectral imaging systems typically employ full-band scanning. In demanding point-of-care testing (POCT) environments such as intensive care units (ICUs), patient agitation or ventilator vibrations can cause long scan times for traditional hyperspectral systems, making them prone to image tearing and spatial misalignment due to motion, thus introducing motion artifacts. This embodiment introduces a hardware-software co-operational intelligent truncation scanning mechanism. Through interpretability analysis of a deep learning model, it guides the front-end mechanical system to perform "on-demand truncation scanning," achieving closed-loop optimization of "algorithm identification of core bands, reverse mapping of physical travel, and execution of truncation scanning," fundamentally eliminating motion artifacts.

[0051] (I) Spectral Dimension Activation Analysis and Target Band Screening Based on Grad-CAM For the non-invasive microcirculation monitoring task in sepsis, the system utilizes pre-acquired continuous hyperspectral data to train a ResNet50 deep learning model for end-to-end direct differentiation of "non-sepsis," "sepsis," and "septic shock." To overcome the black-box nature of deep learning and guide hardware acquisition strategies, the Grad-CAM global activation weights in the spectral dimension of the three-classification model were extracted. ; In the formula,α k c This represents the global attention weight of the k-th feature channel extracted by the classification network for a specific target category c (such as a specific physiological state or a specific target object); L is the sequence length of the feature map of this layer in the spectral dimension. Y c This represents the original prediction score of the classification decision model for the target category c (i.e., the network output before passing through the Softmax activation layer). A i k This represents the activation value of the k-th feature map at position i in the spectral dimension in the last convolutional layer of the model. The partial derivative (i.e., gradient) of the predicted score with respect to the activation value at that location is used to characterize the relative contribution of local spectral features to the final classification decision.

[0052] like Figure 6 As shown, the results indicate that the model's attention is highly focused on the visible light region from 400 nm to 760 nm; in particular, the short-wavelength region from 400 nm to 500 nm exhibits the strongest feature activation. From a pathophysiological optics perspective, this short-wavelength region perfectly covers the strong Soret absorption band of hemoglobin, making it extremely sensitive to early superficial capillary ischemia and microthrombus formation, while also capturing the optical features of jaundice (abnormal bilirubin) that often accompanies sepsis.

[0053] Although the near-infrared band (e.g., greater than 900 nm) contains features reflecting deep tissue edema (tissue water index TWI) caused by capillary leakage, in order to eliminate motion artifacts, the system sets a strategy threshold and actively truncates the long band greater than 760 nm, and finally extracts 400 nm to 760 nm (a total of 180 core channels) as the target feature band range for truncated scanning.

[0054] (ii) Inverse mapping from spectral range to physical control coordinates After obtaining the core spectral range, the main control software calls the wavelength inverse mapping function determined during the system calibration stage. Based on the inverse operation of the formula, it accurately transforms the physical boundaries of the target band (400 nm and 760 nm) into the absolute mechanical coordinates x_start and x_end of the one-dimensional linear displacement stage.

[0055] ; In the formula, x represents the absolute physical position coordinates to which the one-dimensional linear displacement stage needs to be precisely moved after the reverse mapping calculation; λ is the upper and lower boundary wavelengths of the target core spectral band determined by the attention weight distribution. w center1 represents the edge pixel coordinates corresponding to the width dimension of the camera detector target surface (ensuring the wavelength range covers the camera target surface); 2 represents the linear dispersion calibration coefficient of the system along the scanning direction of the displacement stage, reflecting the mapping slope between physical displacement and center wavelength; 3 represents the camera pixel coordinates and wavelength calibration coefficients; 4 represents the initial intercept constant obtained by the system through spectral calibration.

[0056] (III) Hardware Co-execution and High-Speed ​​Inference of Truncation Scanning Action Physical truncation scanning: In actual bedside acquisition, the one-dimensional linear displacement stage is controlled to perform scanning only within the physical effective stroke from x_start to x_end, and the camera is synchronously triggered to acquire data by the position feedback sensor; after scanning to x_end (corresponding to 760 nm), the system immediately terminates acquisition and resets, completely skipping the mechanical push-broom of the background band >760 nm.

[0057] Hardware speedup effect: Redundant bands are directly discarded at the physical level, which greatly reduces the mechanical action time of a single hyperspectral scan from 5 seconds to 3 seconds, achieving a 40% hardware speedup and fundamentally eliminating the interference of most motion artifacts.

[0058] High-speed edge reconstruction and prediction: Based on the frame sequence obtained from the above truncation, the system offline simulates truncation scan reconstruction and performs 5-fold cross-validation. For example... Figure 7 As shown, the cutoff model (400-760 nm) still maintains extremely high prediction performance, with an average accuracy of 91.25% ± 0.56%. Figure 8 As shown, the area under the multivariate ROC curve (AUC) for non-septic, septic, and septic shock reached 0.9642 ± 0.0095, 0.9999 ± 0.0001, and 0.9342+0.0084, respectively.

[0059] In summary, this embodiment achieves a significant physical speedup by sacrificing a very small amount of algorithm-hardware trade-off, perfectly adapting to the real-time reading requirements in harsh environments, and providing invaluable zero-wait support for cardiopulmonary resuscitation and clinical decision-making in intensive care. Example 3: Shortwave Infrared Medicinal Herb Identification Device

[0060] This embodiment relates to a short-wave infrared hyperspectral imaging device for rapid identification of traditional Chinese medicinal materials. A photograph of the prototype in operation is shown below. Figure 12 As shown. The device operates in the short-wave infrared region of 1000-1700 nm, which is rich in characteristic absorption information of organic molecules (such as CH and OH bonds).

[0061] The system's hardware configuration is specifically designed for short-wave infrared characteristics. Imaging objective 1 employs a dedicated SWIR lens that transmits short-wave infrared materials, while camera 6 uses a detector based on indium gallium arsenide (InGaAs) to ensure high sensitivity and responsivity in this band. The spectroscopic element uses a linearly graded filter (LVF), specifically a linearly graded bandpass filter in this embodiment. Its center wavelength changes linearly along the length of the filter, continuously grading from 1000 nm at one end of the stage to 1700 nm at the other, with a gradation rate of 12.7 nm / mm. The full width at half maximum (FWHM) of the spectral window is approximately 15 nm.

[0062] The drive mechanism employs a one-dimensional linear displacement stage 2 driven by a high-precision linear motor to meet the stringent requirements of repeatability and consistency in band positioning for short-wave infrared spectroscopy analysis. In the optical path design, the telecentric imaging lens group 5 plays a crucial role: on the one hand, its special lens coating provides anti-reflection protection for the short-wave infrared band, effectively suppressing stray light; on the other hand, its telecentric structure ensures that the principal ray passes perpendicularly through the linearly graded filter, fundamentally solving the problem of significant center wavelength drift caused by large-angle incident light in the infrared band, thus guaranteeing the authenticity and comparability of the spectral data.

[0063] In the case of Figure 9 When identifying the medicinal materials such as Panax notoginseng and Polygonatum sibiricum, the system workflow is as follows: The displacement stage moves the linear gradient filter at a constant speed, the camera is triggered synchronously, and a series of continuous narrow-band images of the medicinal material surface are quickly acquired, ultimately constructing a complete hyperspectral data cube containing both spatial two-dimensional and spectral dimensions. This data not only records the morphological characteristics of the medicinal material surface (short-wave infrared reflectance spectra under single bands, such as...), but also... Figure 10 As shown in the figure, it also contains information about the chemical composition of its near-surface material.

[0064] By analyzing the spectral curves and absorption peak shapes of the data cube at characteristic wavelengths such as 1450 nm (characteristic absorption of water and hydroxyl groups) and 1650 nm (carbon-hydrogen bond frequency absorption), the absorption peaks were analyzed. Figure 11 As shown, the system can effectively distinguish the authenticity and type of medicinal materials, and even identify the same medicinal material from different origins. This embodiment combines a clever telecentric optical path design with a cost-effective one-dimensional scanning mechanism, which significantly reduces equipment cost and size while ensuring professional-grade spectral performance. This allows the device to be easily deployed at rapid testing stations in places such as medicinal material wholesale markets and pharmaceutical factory quality inspection workshops, providing a powerful technical tool for quality control in the circulation of Chinese medicinal materials.

[0065] The embodiments described above can be further combined or replaced, and these embodiments are merely descriptions of preferred embodiments of the present invention, not limitations on the concept and scope of the present invention. Various changes and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the inventive concept are all within the protection scope of the present invention. The protection scope of the present invention is given by the appended claims and any equivalents.

Claims

1. A smart variable bandwidth hyperspectral imaging device with a linear gradient filter, characterized in that, include: An imaging objective, a graduated filter assembly, a telecentric imaging lens group, and a camera are arranged sequentially along the incident direction of the light path. The gradient filter assembly is mounted on a one-dimensional linear displacement stage, and the displacement stage drives the gradient filter assembly to perform linear scanning motion in a direction perpendicular to the optical axis of the system. The gradient filter assembly has a center wavelength transmittance characteristic that varies linearly along its surface spatial position, and is used to split the incident beam. The telecentric imaging lens group is used to project the split beam onto the photosensitive target surface of the camera to ensure that the main ray is incident parallel to the optical axis and to ensure the consistency of spectral characteristics at each position of the filter. The camera continuously acquires images at a high frame rate during the scanning process, and each frame image corresponds to a scanning position of the gradient filter assembly and the corresponding spectral band. The one-dimensional linear displacement stage is equipped with a position feedback sensor to provide real-time feedback on the current position of the gradient filter assembly, thereby establishing a mapping relationship between the scanning position and the center wavelength. The spectral calibration process for hyperspectral data includes: (i) Remove the imaging objective lens and illuminate the device using a wide-field standard characteristic spectrum light source; (ii) Extract the pixel width coordinate w of the feature spectral line from the image corresponding to the position x of each one-dimensional linear displacement stage, and combine it with the known standard wavelength λ0 to establish the wavelength function relationship λ(x,w) = a·x + b·w + c using the least squares method; (iii) The data reconstruction process includes: calculating the actual center wavelength corresponding to each pixel of each frame image according to the wavelength function λ(x,w); performing spatial alignment and stitching of the image sequence based on the calculation results; setting a fixed spectral coordinate grid, and reconstructing the spectral information of all pixels to the grid through an interpolation algorithm to generate a hyperspectral three-dimensional data cube; The device also includes an intelligent scanning control unit: The intelligent scanning control unit is electrically connected to the one-dimensional linear displacement stage and the camera, and is configured to: acquire a pre-trained attention weight model for a specific classification task; extract the spectral feature attention distribution of the target object based on the model, and identify the core continuous band interval with high contribution and the background band interval with low contribution; The intelligent scanning control unit maps the core band interval and background band interval to the physical start and end coordinates of a one-dimensional linear displacement stage. It performs truncated scanning only within the effective stroke formed by the physical start and end coordinates. Based on the truncated scan frame sequence, it reconstructs a truncated hyperspectral data cube that removes redundant background bands and inputs it into the classification decision model for retraining to complete classification or prediction, thereby improving the acquisition and recognition speed and suppressing motion artifacts.

2. The apparatus as claimed in claim 1, characterized in that, The device also includes a bandwidth adjustment mechanism for continuously fine-tuning the synthesized spectral bandwidth by adjusting the relative position or transmission characteristics of the filters inside the gradient filter assembly. The gradient filter assembly includes a linearly gradient long-pass filter and a linearly gradient short-pass filter stacked in parallel along the optical axis. The two filters change in the same direction and have a peak transmittance of not less than 80%. They form an adjustable bandwidth bandpass window by adjusting their relative positions. The bandwidth adjustment mechanism is used to adjust the relative positions of the two filters to achieve continuous fine-tuning of the bandwidth in the range of 5 nm to 30 nm.

3. The apparatus as described in claim 1, characterized in that, The gradient filter assembly is a single linear gradient bandpass filter with a transmission center wavelength that changes linearly and continuously along one geometric axis of the filter. The wavelength gradient range is 400 nm to 1000 nm, the full width at half maximum (FWHM) is 6 to 20 nm, and the peak transmittance is not less than 80%.

4. The apparatus as claimed in claim 1, characterized in that, The telecentric imaging lens group is an object-oriented telecentric lens or a double telecentric lens, whose object plane position matches the center position of the graduated filter assembly to ensure that the incident principal ray is parallel to the optical axis, and the optical system error of the lens group is controlled within λ / 10.

5. The apparatus as claimed in claim 1, characterized in that, The camera is a high frame rate area array detector, which is electrically connected to a one-dimensional linear displacement stage through a synchronization control unit. During the scanning process, the camera triggers acquisition at a preset frame rate to achieve precise synchronization between image acquisition and filter position. The one-dimensional linear displacement stage includes a precision electric slide and a position feedback sensor. The sensor is used to provide real-time feedback on the current position of the gradient filter assembly and to support wavelength positioning and data reconstruction.

6. The apparatus as claimed in claim 2, characterized in that, The bandwidth adjustment mechanism achieves fine-tuning of the synthesized spectral bandwidth by adjusting the relative position or transmission characteristics of each filter inside the gradient filter assembly. The adjustment range is from 5 nm to 30 nm, and the adjustment step is ≤10 μm.

7. The apparatus as claimed in claim 1, characterized in that, The process of establishing the wavelength function λ(x,w) also includes: performing multi-frame averaging on the calibration image sequence to reduce noise and improve fitting accuracy, with a fitting residual ≤2 nm.

8. The apparatus as claimed in claim 1, characterized in that, The device is suitable for applications such as microcirculation monitoring in intensive care units (ICUs) for sepsis and septic shock, tissue blood oxygenation imaging, shortwave infrared identification of medicinal materials, agricultural remote sensing, food safety testing, or industrial sorting.

9. A method for acquiring hyperspectral data using the device as described in claim 1 or 7, characterized in that, Includes the following steps: S1: The light from the target scene is introduced into the optical path through the imaging objective lens; S2: Control the one-dimensional linear displacement stage to drive the gradient filter assembly to perform linear scanning in the direction perpendicular to the optical axis. The speed is dynamically matched according to the 60 fps frame rate and 2 nm step size to achieve full spectrum acquisition in 5 seconds or acquisition of a custom truncation spectral range. S3: During the scanning process, the camera synchronously acquires a series of two-dimensional narrowband images through the filter at a preset frame rate ≥60 fps. Each frame corresponds to the scanning position x and the center wavelength λ. S4: Record the scanning position x corresponding to each frame of the image, and determine the center wavelength corresponding to each frame of the image based on the linear dispersion characteristics of the gradient filter and the wavelength function λ(x,w). S5: Spatial alignment and spectral interpolation reconstruction are performed on the acquired image sequence to generate a hyperspectral three-dimensional data cube containing spatial and spectral information. The non-uniformly distributed spectral information is mapped to a standard grid with a step size of 2 nm using a cubic spline interpolation algorithm.

10. The method as described in claim 9, characterized in that, Step S2 further includes an adaptive truncation scan control process based on an AI attention mechanism: S2-1: Intelligent division of core band intervals: Using a pre-trained classification decision model combined with an interpretability algorithm, the attention weight distribution of the model in the full spectrum dimension is extracted; based on the set weight threshold, one or more core continuous spectral band intervals that contribute the most to the decision of the current classification task are automatically determined. S2-2: Control parameter reverse physical mapping: Using the wavelength function λ(x,w)= a·x + b·w + c described in the spectral calibration process of hyperspectral data in claim 1, the wavelength boundary of the core continuous spectral band interval is directly reverse mapped to the physical start position coordinates and end position coordinates of a one-dimensional linear displacement stage. S2-3: Truncation Scan Execution and Data Reconstruction: The one-dimensional linear displacement stage is controlled to directly position itself to the physical starting position coordinates, and truncated scanning is performed only within the effective travel range formed by the physical starting position coordinates and the ending position coordinates; the camera is triggered by the position feedback of the one-dimensional linear displacement stage during the truncated scanning and synchronously acquires the corresponding frame sequence; acquisition stops after scanning to the ending position coordinates, and scanning of non-core band intervals is completely skipped; based on the truncated scan frame sequence, a truncated hyperspectral data cube with redundant background bands removed is reconstructed and input into the classification decision model to complete classification or prediction.

Images