A non-reconstruction measurement field target recognition method, system and device based on a near-infrared perovskite detector

Abstract

The application provides a non-reconstruction measurement domain target recognition method, system and equipment based on a near-infrared perovskite detector, the target recognition method comprising: building a single-pixel photoelectric hybrid neural network; pre-training the neural network using a sample set, fixing the network parameters after training is completed, and generating a target reference feature library; forming spectral-spatial hybrid coding and an original measurement value sequence by optically irradiating and modulating the light path of a target to be recognized; pre-processing the original measurement value sequence to obtain a standard measurement value sequence; inputting the obtained standard measurement value sequence into the pre-trained neural network for feature extraction and target matching recognition, and outputting a target recognition result. The application realizes direct mapping from a measurement value sequence to a target category through the collaborative design of a near-infrared perovskite detector and a non-reconstruction measurement domain recognition task, and the target recognition can be completed without performing an image reconstruction process throughout the whole process. The application is suitable for the field of perovskite optical imaging technology.

Description

Technical Field

[0001] This application relates to the field of perovskite optical imaging technology, specifically to a non-reconstruction measurement domain target recognition method, system, and device based on a near-infrared perovskite detector. Background Technology

[0002] The rapid development of photoelectric detection technology and its breakthroughs in stability and noise control have led to continuous improvements in the performance of single-pixel detectors. Currently, mainstream near-infrared detectors are mainly based on semiconductor materials such as silicon, germanium, and indium gallium arsenide. Among them, silicon-based detectors have limited response capabilities in the near-infrared band, while detectors such as germanium and indium gallium arsenide have better near-infrared response performance. However, due to issues such as complex manufacturing processes and high costs, high-performance near-infrared detectors have long been limited to certain fields.

[0003] Perovskite materials, with their continuously tunable bandgap within a certain range, high absorption coefficient, and high carrier mobility, have become an ideal choice for fabricating high-performance photodetectors. Especially in the near-infrared band, perovskite single-crystal thin films, due to their low defect density and excellent photoelectric properties, significantly improve the sensitivity and response speed of detectors. Furthermore, perovskite materials can be prepared using solution methods at low temperatures, resulting in simple and low-cost processes that are compatible with flexible substrates, providing an ideal material choice for fabricating high-performance, low-cost near-infrared single-pixel detectors.

[0004] However, existing detector technologies still face systemic challenges in achieving near-infrared target recognition. Traditional target recognition methods typically rely on complex signal reconstruction processes, requiring image reconstruction before recognition. Furthermore, single-point devices only output a one-dimensional total light intensity value, lacking the parallel readout capability required for area array imaging. Forcing image reconstruction necessitates complex iterative optimization algorithms, which not only amplifies the low-frequency drift noise of perovskite detectors under low-light conditions but also weakens the high signal-to-noise ratio advantage of perovskite due to error accumulation during iteration. This reconstruction process not only increases computational complexity but also significantly impacts the quality of the reconstructed image, easily leading to target feature loss under low sampling rates and low-light conditions, thus affecting subsequent recognition accuracy. Consequently, the inherent advantages of perovskite detectors, such as high sensitivity and low noise, cannot be fully realized in the final recognition results.

[0005] Therefore, existing technologies lack a target recognition method that organically combines near-infrared perovskite detectors, single-pixel measurement systems, and non-reconstruction measurement domain target recognition strategies; otherwise, it will be difficult to fully leverage the application advantages of perovskite detectors in near-infrared target recognition scenarios. Summary of the Invention

[0006] To address the aforementioned technical deficiencies, this application provides a method, system, and device for non-reconstruction measurement domain target recognition based on a near-infrared perovskite detector.

[0007] According to the first aspect of this application, a non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector is provided, comprising: A single-pixel optoelectronic hybrid neural network based on a near-infrared perovskite detector and having both optical and electrical end networks was constructed. Construct a target recognition sample set, wherein the sample set data includes similar sample pairs and dissimilar sample pairs; A single-pixel optoelectronic hybrid neural network was pre-trained using a target recognition sample set. After training, the network parameters were fixed and a target reference feature library was generated. By illuminating the target with optical light and modulating the optical path, a spectral-spatial hybrid code and a sequence of original measurement values ​​are formed. The original measurement value sequence is preprocessed based on the detector response model to obtain the standard measurement value sequence; The obtained standard measurement value sequence is input into a pre-trained single-pixel photoelectric hybrid neural network for feature extraction and target matching and recognition, and the target recognition result is output. The optical network is a coding layer used to achieve spectral-spatial hybrid coding; the electrical network is a decoding layer, including a preprocessing layer, a feature extraction layer, and a matching and recognition layer, used to identify targets from single-pixel detection signals. The optical illumination source is a near-infrared broadband light source; The target recognition process does not require the generation of a target reconstruction image throughout.

[0008] Preferably, the step of forming a spectral-spatial hybrid code and a sequence of original measurement values ​​by optically illuminating and modulating the target to be identified specifically includes: A broadband light beam is emitted using a near-infrared broadband light source. After passing through a lens group, the light beam converges to form an image after passing through the target to be identified. The transmitted beam carrying the target's spatial information is dispersed in space according to wavelength by a dispersive element to form a one-dimensional wavelength distribution. By using a focusing lens, all wavelength components after dispersion are converged onto the surface of the digital micromirror device to form a spectral-spatial hybrid coding pattern; A digital micromirror device is used to load a digital coded aperture to spatially modulate the intensity of a hybrid coded pattern and reflect it to a perovskite single-point detector. The received optical signal is converted into an electrical signal by a perovskite point detector, and the original measurement value sequence corresponding to the coded pattern is obtained.

[0009] Preferably, the step of using a digital micromirror device to load a digital coded aperture to spatially modulate the intensity of the hybrid coded pattern and reflect it to a perovskite single-point detector specifically includes: The digital micromirror device uses a complementary coding pattern generated by a Hadamard matrix for cutting a cake. It is divided into two groups, positive and negative, with the matrix values ​​of the two groups being opposite and complementary. The digital micromirror device is loaded with positive and negative complementary coded apertures in a predetermined order, and the near-infrared perovskite single-point detector sequentially collects the corresponding positive and negative measurement values. In this device, each row and column of pixels on the surface of the digital micromirror device corresponds to a specific wavelength range and spatial position. While loading the coded aperture, each micromirror unit synchronously modulates the spectral components of its corresponding spatial position independently, coupling the spatial position and wavelength information into a one-dimensional optical signal.

[0010] Preferably, the preprocessing of the original measurement value sequence based on the detector response model to obtain the standard measurement value sequence specifically includes: The data acquisition unit receives the analog electrical signal output by the perovskite point detector and converts it into a digital signal; The data processing algorithm corrects the acquired measurement sequence based on the physical response parameters of the perovskite detector, including dark current compensation, responsivity compensation, and normalization. The physical response parameters of the perovskite detector include dark current, noise level, quantum efficiency, responsivity, and near-infrared band response characteristics. Background subtraction is performed on the original measurement signal based on the detector's dark current parameters to obtain the measurement value sequence after eliminating background noise: M1 = M raw - I dark M raw For the original measurement sequence, I dark This is the dark current signal of the detector; Based on the measured response characteristics of the detector, a correction coefficient K is constructed to compensate for the measured values, resulting in the compensated measured value sequence: M2=M1 K; The standard measurement value sequence is obtained by normalizing the detector noise characteristics: M3=(M2-μ) / σ, where μ is the mean of the measurement signal and σ is the standard deviation of the measurement signal.

[0011] Preferably, the step of inputting the obtained standard measurement value sequence into a pre-trained single-pixel photoelectric hybrid neural network for feature extraction and target matching recognition, and outputting the target recognition result, specifically includes: The obtained standard measurement value sequence is input into a pre-trained single-pixel photoelectric hybrid neural network for feature extraction, so that the high-dimensional measurement value sequence is mapped into a low-dimensional feature vector. The low-dimensional feature vector is matched with the reference feature vector in the target reference feature library, the Euclidean distance is calculated, and the category with the smallest distance is output as the target recognition result. The Euclidean distance D is calculated using the Euclidean distance algorithm: ;

[0012] in, represents the squared distance between the two input samples after being mapped by the neural network, and h represents the output vector of the hidden layer of the neural network. This represents the square of the L2 norm obtained by outputting the same network twice.

[0013] Preferably, the calculation of the Euclidean distance, outputting the category with the smallest distance as the target recognition result, specifically includes: When the calculated minimum Euclidean distance is less than a preset threshold, the target recognition result corresponding to the minimum Euclidean distance is output; when the calculated minimum Euclidean distance is greater than or equal to the preset threshold, the target to be identified is determined to be an unknown target, an unknown target signal is sent to the system and retraining extension is triggered. After receiving an unknown target signal, the system assigns a new category identifier to the unknown target, binds the low-dimensional feature vector of the unknown target to the category identifier, and stores it as a new reference feature vector in the target reference feature library, thus completing the triggered retraining-free extension. During the retraining-free extension process, the network parameters remain frozen.

[0014] Preferably, the step of constructing a target recognition sample set includes similar sample pairs and dissimilar sample pairs; the target recognition sample set is used to pre-train a single-pixel photoelectric hybrid neural network, and after training, the network parameters are fixed and a target reference feature library is generated; specifically, this includes: The single-pixel optoelectronic hybrid neural network uses a twin network for feature extraction. The twin network consists of two encoders with identical structures and shared weight parameters. The two encoders receive two sets of measurement value sequences as input and output corresponding feature vectors. The pre-training process includes: The constructed target recognition sample set includes a training set and a validation set. The same type of sample pairs and different type of sample pairs in the training set are used as training data and input into a single-pixel photoelectric hybrid neural network as the target to be recognized. After optical illumination and optical path modulation, a spectral-spatial hybrid code is formed. By calculating the differential measurement values ​​and preprocessing the detector response model, a standard measurement value sequence is obtained and input into the twin network. In the Siamese network, the two encoders perform forward propagation on the two sets of input sequences through a multi-layer nonlinear mapping structure, reducing the original high-dimensional measurement values ​​to low-dimensional feature vectors. By continuously optimizing the network parameters through the comparative loss function until the loss function converges or reaches the preset maximum number of iterations, the network parameters are fixed, and the low-dimensional feature vectors of the samples are extracted using the network to construct the target reference feature library, thus completing the pre-training.

[0015] Preferably, the contrast loss function is: L = y·D² + (1-y)·max(margin - D, 0)²; Where y is the label of the sample pair, with a value of 1 for similar sample pairs and a value of 0 for dissimilar sample pairs; D is the Euclidean distance between feature vectors; margin is a preset threshold parameter used to control the minimum interval of dissimilar sample pairs, with a value of 0.8.

[0016] According to a second aspect of this application, a non-reconstruction measurement domain target recognition system based on a near-infrared perovskite detector is provided, comprising: The optical illumination and modulation module includes a near-infrared broadband light source, a beam expander / collimator lens group, an imaging lens, a dispersive element, a focusing lens, and a digital micromirror device, used to generate spectral-spatial hybrid coded patterns and perform spatial light intensity modulation; The photoelectric detection module includes a near-infrared perovskite single-point detector and a data acquisition unit, which is used to acquire modulated optical signals and generate a sequence of raw measurement values. The data processing and recognition module includes a data processing unit, which is used to preprocess the original measurement value sequence and reduce its dimension to a low-dimensional feature vector, calculate the minimum matching distance, send unknown target signals, and output target recognition results; The rejection and expansion module is used to receive and respond to unknown target signals, assign new category identifiers to unknown targets, and store the low-dimensional feature vectors of unknown targets into the target reference feature library.

[0017] According to a third aspect of this application, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector.

[0018] The target recognition method provided in this application achieves a direct mapping from the measurement value sequence to the target category through the collaborative design of a near-infrared perovskite detector and a non-reconstruction measurement domain recognition task. This allows the target recognition task to be completed without performing an image reconstruction process. Utilizing a near-infrared broadband light source in conjunction with a dispersive element, joint encoding and preprocessing of spatial and spectral information are implemented at the optical front end. This enables the light signal modulated by the digital micromirror device to carry richer target category features. Simultaneously, a detector response model is established using the dark current, responsivity, quantum efficiency, and near-infrared response characteristics of the perovskite detector, and the measurement value sequence is physically corrected and its features enhanced. Then, based on the corrected measurement value sequence, measurement domain feature extraction and target recognition are performed, achieving synergistic optimization of the detector's physical characteristics and the recognition algorithm.

[0019] To address the problems of excessive reliance on image reconstruction, high computational complexity, and limited recognition efficiency under low-light conditions in near-infrared target recognition using existing detector technologies, this application proposes a non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector. Compared to existing methods, the method in this application does not simply replace the detector, but rather integrates the design from the optical encoding end to the signal detection end and then to the recognition decision end. Through a spectral-spatial hybrid encoding mechanism, the measurement value sequence is perfectly adapted to the neural network recognition task, allowing the performance advantages of the near-infrared perovskite single-point detector to be directly translated into advantages in target recognition performance. While ensuring recognition accuracy, the image reconstruction process is avoided, providing a highly efficient, low-complexity, and scalable technical solution for near-infrared target recognition.

[0020] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of what is pointed out in the written description and the accompanying drawings. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart of the non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector provided in this application.

[0022] Figure 2 This is a schematic diagram of the non-reconstruction measurement domain target recognition system based on a near-infrared perovskite detector provided in this application.

[0023] Figure 3A schematic diagram of the feature extraction process of the Siamese network based on the pre-training process provided in this application.

[0024] Figure 4 The confusion matrix provided for the accuracy of Chinese character recognition in this application.

[0025] Figure 5 Complementary digital coded aperture provided for this application. Detailed Implementation

[0026] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0027] To address some problems existing in the current technology: In a first aspect, this application provides a non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector.

[0028] For example, this non-reconstruction measurement domain target recognition method based on near-infrared perovskite detectors, such as Figure 1 As shown, the non-reconstruction measurement domain target recognition method includes: A single-pixel optoelectronic hybrid neural network based on a near-infrared perovskite detector and having both optical and electrical end networks was constructed. Construct a target recognition sample set, wherein the sample set data includes similar sample pairs and dissimilar sample pairs; A single-pixel optoelectronic hybrid neural network was pre-trained using a target recognition sample set. After training, the network parameters were fixed and a target reference feature library was generated. By illuminating the target with optical light and modulating the optical path, a spectral-spatial hybrid code and a sequence of original measurement values ​​are formed. The original measurement value sequence is preprocessed based on the detector response model to obtain the standard measurement value sequence; The obtained standard measurement value sequence is input into a pre-trained single-pixel photoelectric hybrid neural network for feature extraction and target matching and recognition, and the target recognition result is output. The optical network is a coding layer used to achieve spectral-spatial hybrid coding; the electrical network is a decoding layer, including a preprocessing layer, a feature extraction layer, and a matching and recognition layer, used to identify targets from single-pixel detection signals. The optical illumination source is a near-infrared broadband light source; The target recognition process does not require the generation of a target reconstruction image throughout.

[0029] Based on the above scheme, the method provided in this application organically combines a near-infrared perovskite detector, a single-pixel measurement system, and a non-reconstruction measurement domain target recognition strategy. With the perovskite detector as the core, high-precision recognition of measurement domain signals is achieved through integrated collaborative design. This not only fully leverages the application advantages of perovskite detectors in near-infrared target recognition scenarios but also maintains good scalability and transferability. It can directly utilize the measurement signals collected by the detector to complete feature extraction and target recognition, fully leveraging the performance advantages of perovskite detectors while avoiding the image reconstruction process, thereby improving recognition efficiency and system applicability.

[0030] In some possible implementations of the first aspect, a spectral-spatial hybrid code and a sequence of original measurements are formed by optically illuminating and modulating the target to be identified, specifically including: A broadband light beam is emitted using a near-infrared broadband light source. After passing through a lens group, the light beam converges to form an image after passing through the target to be identified. The transmitted beam carrying the target's spatial information is dispersed in space according to wavelength by a dispersive element to form a one-dimensional wavelength distribution. By using a focusing lens, all wavelength components after dispersion are converged onto the surface of the digital micromirror device to form a spectral-spatial hybrid coding pattern; A digital micromirror device is used to load a digital coded aperture to spatially modulate the intensity of a hybrid coded pattern and reflect it to a perovskite single-point detector. The received optical signal is converted into an electrical signal by a perovskite point detector, and the original measurement value sequence corresponding to the coded pattern is obtained.

[0031] Based on the above scheme, the perovskite single-point detector is a photosensitive non-array device made of perovskite material. The selected perovskite layer is two-dimensional or three-dimensional perovskite, and it is prepared using tin-lead mixed perovskite material. This gives the perovskite single-point detector high near-infrared responsivity, high signal-to-noise ratio, and low dark current characteristics. Its spectral response range matches the operating wavelength of broadband light sources and dispersive elements (780nm~1100nm), and it has the photoelectric conversion capability to convert optical signals into electrical signals. Therefore, compared with traditional silicon-based detectors, perovskite detectors can obtain higher quality measurement signals in the near-infrared band, providing a reliable data foundation for subsequent acquisition of high-quality measurement value sequences of the total light field intensity after spectral-spatial hybrid encoding and for target recognition in the measurement domain.

[0032] The relative positions of the imaging lens, focusing lens, and digital micromirror device (DMD) need to be optimized based on the photosensitive area and receiving field of view of the perovskite point detector to ensure that the reflected light field modulated by the digital micromirror device falls completely within the detector range. The dispersive element is any one of a transmission grating, a reflection grating, or a dispersive prism, and its operating wavelength covers 780nm to 1100nm, matching the output wavelength of the near-infrared broadband light source.

[0033] The dispersive element is disposed between the imaging lens and the focusing lens to disperse the broadband beam after it has been converged by the imaging lens in a preset direction in space according to the wavelength, so that the light components of different wavelengths form a one-dimensional wavelength distribution in space.

[0034] A focusing lens is positioned after the dispersive element to completely converge all wavelength components after dispersion to the surface of the digital micromirror device, so that different spatial positions in the row or column direction of the digital micromirror device surface correspond to different wavelength ranges, thereby realizing the mixed encoding of target spatial information and spectral information.

[0035] The angular dispersion or linear dispersion of the dispersive element is matched with the pixel size of the digital micromirror device, so that the spatial separation distance between adjacent wavelength components on the surface of the digital micromirror device is not less than the width of a single pixel and not greater than the preset field of view width, so as to ensure that each wavelength component can be independently modulated by different pixel columns or rows of the digital micromirror device.

[0036] Specifically, the dispersive element unfolds the spectral information of different spatial locations of the target on the DMD surface in an ordered manner according to wavelength, so that each column (or row) of pixels in the DMD corresponds to a specific wavelength range, and each row (or column) of pixels corresponds to a different spatial location of the target. When the DMD is loaded with a digitally coded aperture, each micromirror unit simultaneously modulates the spectral component at its corresponding position independently, so that a single measurement value is not only a projection of the target's spatial information, but also a weighted superposition of spatial and spectral information. This design embeds spectral dimension information, which originally required multiple spectral scans or additional spectrometers to acquire, into the single-pixel measurement value sequence with zero additional measurement times, significantly enriching the information dimension of the measurement value sequence without increasing the sampling burden. In addition, the spectral-spatial hybrid coding mechanism is naturally adaptable to subsequent neural network recognition tasks. Since different target categories often have different spectral reflectance and transmission characteristics in the near-infrared band, after dispersive unfolding, these differences in spectral characteristics will be encoded as differences in light intensity distribution on different pixel columns of the digital micromirror device, and thus reflected in different positional intervals of the measurement value sequence. This allows the neural network to directly perceive target information from a one-dimensional measurement value sequence without complex feature engineering.

[0037] Meanwhile, the redundant spectral information introduced by hybrid coding can effectively alleviate the over-reliance of single spatial coding on target texture and edge features, enhance the ability of the measurement sequence to represent the differences in target categories, and enable the subsequent neural network to still obtain high classification accuracy under low sampling rate conditions, and have stronger robustness to illumination fluctuations and environmental noise.

[0038] In the actual sampling process, the measurement values ​​of each coded aperture are collected in positive and negative complementary modes and differential operations are performed. This can cancel the common-mode low-frequency noise and dark current drift of the perovskite detector within a certain period of time, so that the differential measurement value sequence is only related to the spectral-spatial mixed information of the target.

[0039] In some possible ways of implementing the first aspect, such as Figure 5 As shown, the method of using a digital micromirror device to load a digital coded aperture to spatially modulate the intensity of the hybrid coded pattern and reflect it to a perovskite single-point detector specifically includes: The digital micromirror device uses a complementary coding pattern generated by a Hadamard matrix for cutting a cake. It is divided into two groups, positive and negative, with the matrix values ​​of the two groups being opposite and complementary. The digital micromirror device is loaded with positive and negative complementary coded apertures in a predetermined order, and the near-infrared perovskite single-point detector sequentially collects the corresponding positive and negative measurement values. In this device, each row and column of pixels on the surface of the digital micromirror device corresponds to a specific wavelength range and spatial position. While loading the coded aperture, each micromirror unit synchronously modulates the spectral components of its corresponding spatial position independently, coupling the spatial position and wavelength information into a one-dimensional optical signal.

[0040] Based on the above scheme, the refresh frequency of the complementary digital coded aperture loaded by the digital micromirror device (DMM) matches the response time constant of the perovskite detector. Each time the DMM completes loading a coded aperture, it sends a corresponding trigger signal to the data acquisition module to record the current coded aperture number. Since the surface of the DMM carries a spectral-spatial mixed pattern after dispersion expansion, each column (or row) of pixels in the DMM corresponds to a specific wavelength range. Therefore, when loading a coded aperture, each micromirror unit independently modulates the spectral component at its corresponding spatial position, enabling a single measurement value to simultaneously encode multiple feature dimensions of the target. The sampling trigger delay of the data acquisition unit is dynamically calibrated based on the photogenerated carrier mobility and rise / fall times of the perovskite detector to ensure that the amplitude sampling points of the measurement value sequence are within the steady-state range of the perovskite detector's response curve, avoiding amplitude attenuation caused by response hysteresis. Figure 5As shown, the image illustrates the smallest functional unit of a complementary coded aperture (CFA), precisely demonstrating the most fundamental and core concept of "complementary pairs" in CFA. Each small square in the image corresponds to a micromirror on a digital micromirror device (DMM). In practical applications, the DMM displays the left image first, then the right image, within an extremely short time (microseconds). The left image is responsible for acquiring the "positive signal," and the right image is responsible for acquiring the "negative signal," then performing a differential operation. Since the environmental and background noise remain essentially unchanged in the two acquisitions, they are canceled out after subtraction, thus greatly improving the image contrast and signal-to-noise ratio. The black squares represent "light-transmitting" or "reflecting" states, while the white background represents "light-blocking" or "absorption" states. The two are arranged in an alternating distribution; the patterns on the left and right sides exhibit strict inversion complementarity: a black position on the left corresponds to a white position on the right, and vice versa.

[0041] The specific operation is as follows: Digital micromirror devices (DMMs) sequentially load positive and negative complementary coded apertures in a predetermined order. Each time a pattern is loaded, a synchronous trigger signal is sent to the data acquisition unit. Simultaneously, a perovskite single-point detector acquires the reflected light signal modulated by that pattern, obtaining the original positive and negative measurement values. Since the surface of the DMM carries a spectral-spatial hybrid pattern, each column (or row) of pixels corresponds to a specific wavelength range. Therefore, each time a coded aperture is loaded, each micromirror unit independently modulates the spectral component at its corresponding spatial position, enabling a single measurement value to simultaneously encode multiple feature dimensions of the target. The final differential measurement value effectively suppresses background light and low-frequency detector noise. Because the perovskite detector used has extremely high responsivity and signal-to-noise ratio in the near-infrared band and perfectly matches the operating wavelength (780nm~1100nm) of the broadband light source and dispersive element, high-quality differential signals can be obtained in a single acquisition. The data acquisition unit associates each differential measurement value with the corresponding coded pattern number, forming a sequence of original measurement values ​​with a length equal to the total number of coded patterns. Compared with conventional single-pixel systems that only encode pure spatial information, this application introduces a combination of "broadband light source + dispersive element" so that the measurement value sequence not only contains the spatial projection information of the target, but also contains the spectral response features corresponding to different spatial locations. This significantly enriches the information dimension of the measurement value sequence and provides more sufficient feature basis for subsequent preprocessing and neural network recognition.

[0042] In some possible implementations of the first aspect, the preprocessing of the original measurement sequence based on the detector response model to obtain a standard measurement sequence specifically includes: The data acquisition unit receives the analog electrical signal output by the perovskite point detector and converts it into a digital signal; The data processing algorithm corrects the acquired measurement sequence based on the physical response parameters of the perovskite detector, including dark current compensation, responsivity compensation, and normalization. The physical response parameters of the perovskite detector include dark current, noise level, quantum efficiency, responsivity, and near-infrared band response characteristics. Background subtraction is performed on the original measurement signal based on the detector's dark current parameters to obtain the measurement value sequence after eliminating background noise: M1 = M raw - I dark M raw For the original measurement sequence, I dark This is the dark current signal of the detector; Based on the measured response characteristics of the detector, a correction coefficient K is constructed to compensate for the measured values, resulting in the compensated measured value sequence: M2=M1 K; The standard measurement value sequence is obtained by normalizing the detector noise characteristics: M3=(M2-μ) / σ, where μ is the mean of the measurement signal and σ is the standard deviation of the measurement signal.

[0043] Based on the above scheme, the perovskite detector generates a weak dark current signal under no-light conditions. By subtracting the dark current background signal, the influence of environmental noise can be reduced. The dark current value is subtracted from each element in the obtained differential measurement sequence. The dark current value is a stable value obtained by averaging the detector output multiple times under the condition that the light source is turned off and the optical path is blocked before the system starts working.

[0044] Although perovskite detectors exhibit linear responses over a wide dynamic range, slight nonlinearities may exist at the edges of strong light or near the cutoff band. The detector's responsivity curve R and the detector's response at the current actual incident light power were obtained through pre-testing. The response R(p)f is given, and then a correction coefficient K=R / R(p) is constructed to compensate for the measured value: M2=M1 K corrects response deviations under different operating conditions through responsiveness compensation.

[0045] The noise level of perovskite detectors may vary with signal strength and frequency. To eliminate the interference of noise on feature extraction and make the measurement value sequences of different batches comparable, the compensated measurement value M2 is normalized: M3 = (M2 - μ) / σ, where μ is the mean of the measurement value M2 and σ is the standard deviation. This operation converts the measurement value sequence into a standard normal distribution with a mean of 0 and a variance of 1, eliminating the difference in measurement value amplitude, so as to obtain a feature enhancement signal suitable for direct recognition in the measurement domain, which greatly improves the convergence speed and robustness of the feature extraction network.

[0046] In some possible implementations of the first aspect, the step of inputting the obtained standard measurement value sequence into a pre-trained single-pixel photoelectric hybrid neural network for feature extraction and target matching recognition, and outputting the target recognition result, specifically includes: The obtained standard measurement value sequence is input into a pre-trained single-pixel photoelectric hybrid neural network for feature extraction, so that the high-dimensional measurement value sequence is mapped into a low-dimensional feature vector. The low-dimensional feature vector is matched with the reference feature vector in the target reference feature library, the Euclidean distance is calculated, and the category with the smallest distance is output as the target recognition result. The Euclidean distance D is calculated using the Euclidean distance algorithm: ;

[0047] in, represents the squared distance between the two input samples after being mapped by the neural network, and h represents the output vector of the hidden layer of the neural network. This represents the square of the L2 norm obtained by outputting the same network twice.

[0048] Based on the above scheme, since the signal-to-noise ratio of a single measurement by a perovskite detector is significantly higher than that of a traditional detector in the near-infrared band, the redundancy of target information in the measurement sequence is relatively low. This allows the feature extraction network to avoid designing complex deep convolutional structures and instead adopt a lightweight fully connected structure of no more than 3 layers. This compresses the computational delay while ensuring recognition accuracy and fully leverages the advantage of the perovskite's fast response speed.

[0049] The collaborative design process for target recognition tasks includes the comprehensive integration of the physical characteristics of perovskite detectors, as well as the physical characteristics of the perovskite detector in the process of converting optical signals into electrical signals, such as quantum efficiency, carrier mobility, response time, and pixel size. It also includes noise characteristics constructed based on the measured parameters of the detector, such as dark current noise and response noise. Furthermore, it involves the relationship between the detector's input signal strength and output electrical signal, fully reproducing the physical process of the detector converting near-infrared light signals into measurable electrical signals.

[0050] The relationship between perovskite detectors and identification tasks is not a simple hardware replacement, but rather a synergistic one. Because perovskite detectors have high responsivity and low noise levels in the near-infrared band, they can obtain stable measurement sequences with high signal-to-noise ratios, thereby improving the feature representation capability of the measurement domain and enabling the effective preservation of target information under conditions of fewer samples.

[0051] The feature extraction network establishes a mapping relationship between the measurement value sequence and the target category features, converting the high-dimensional measurement value sequence into a low-dimensional discriminative feature vector. Through network learning, different categories of targets can form a distinguishable distribution structure in the feature space, which is used to represent the measurement domain features of the preprocessed measurement value sequence, mapping the high-dimensional measurement value sequence into a low-dimensional feature vector with category discrimination capability to characterize the feature information of the target to be measured. The feature extraction network is a lightweight mapping network, and its number of network layers and hidden layer dimensions are pre-calibrated and determined based on the signal-to-noise ratio and dynamic range of the measured output signal of the perovskite detector. It constructs a feature mapping relationship for the measurement value sequence output by the near-infrared perovskite detector and uses a neural network model to learn the target feature information in the measurement value sequence, realizing automatic extraction of measurement domain features. The feature extraction network preferably adopts a Siamese network structure, but other models can also be used.

[0052] In some possible implementations of the first aspect, the calculation of the Euclidean distance, outputting the category with the smallest distance as the target recognition result, specifically includes: When the calculated minimum Euclidean distance is less than a preset threshold, the target recognition result corresponding to the minimum Euclidean distance is output; when the calculated minimum Euclidean distance is greater than or equal to the preset threshold, the target to be identified is determined to be an unknown target, an unknown target signal is sent to the system and retraining extension is triggered. After receiving an unknown target signal, the system assigns a new category identifier to the unknown target, binds the low-dimensional feature vector of the unknown target to the category identifier, and stores it as a new reference feature vector in the target reference feature library, thus completing the triggered retraining-free extension. During the retraining-free extension process, the network parameters remain frozen.

[0053] Based on the above scheme, when the data processing and identification module determines that the target to be identified is an unknown target, it sends an unknown target signal to the rejection and expansion module and triggers the retraining-free expansion; after receiving the unknown target signal, the rejection and expansion module assigns a new category identifier to the unknown target, binds the low-dimensional feature vector of the unknown target with the category identifier, and stores it as a new reference feature vector in the target reference feature library, thus completing the triggered retraining-free expansion. During the entire retraining-free expansion process, the network parameters remain frozen.

[0054] This solution combines Siamese network feature learning with a reference feature library matching and recognition mechanism to achieve a non-reconstruction measurement domain target recognition method that can expand categories without image reconstruction or retraining the feature extraction network. Through the cooperation of the data processing and recognition module and the rejection and expansion module, it provides a technical foundation for subsequent applications such as character recognition, target detection, and target classification in complex scenes, thereby improving the scalability and application flexibility of the system.

[0055] In some possible ways of implementing the first aspect, such as Figure 3 As shown, the construction of the target recognition sample set includes similar sample pairs and dissimilar sample pairs; the single-pixel photoelectric hybrid neural network is pre-trained using the target recognition sample set, and after training, the network parameters are fixed and a target reference feature library is generated; specifically, it includes: The single-pixel optoelectronic hybrid neural network uses a twin network for feature extraction. The twin network consists of two encoders with identical structures and shared weight parameters. The two encoders receive two sets of measurement value sequences as input and output corresponding feature vectors. The pre-training process includes: The constructed target recognition sample set includes a training set and a validation set. The same type of sample pairs and different type of sample pairs in the training set are used as training data and input into a single-pixel photoelectric hybrid neural network as the target to be recognized. After optical illumination and optical path modulation, a spectral-spatial hybrid code is formed. By calculating the differential measurement values ​​and preprocessing the detector response model, a standard measurement value sequence is obtained and input into the twin network. In the Siamese network, the two encoders perform forward propagation on the two sets of input sequences through a multi-layer nonlinear mapping structure, reducing the original high-dimensional measurement values ​​to low-dimensional feature vectors. By continuously optimizing the network parameters through the comparative loss function until the loss function converges or reaches the preset maximum number of iterations, the network parameters are fixed, and the low-dimensional feature vectors of the samples are extracted using the network to construct the target reference feature library, thus completing the pre-training.

[0056] In some possible implementations of the first aspect, the contrastive loss function is: L = y·D² + (1-y)·max(margin - D, 0)²; Where y is the label of the sample pair, with a value of 1 for similar sample pairs and a value of 0 for dissimilar sample pairs; D is the Euclidean distance between feature vectors; margin is a preset threshold parameter used to control the minimum interval of dissimilar sample pairs, with a value of 0.8.

[0057] Based on the above scheme, by optimizing the loss function, the feature extraction network learns a discriminative feature space. In this feature space, measurement sequences belonging to the same category are mapped to feature vectors that are close to each other, while measurement sequences from different categories are mapped to feature vectors that are far apart from each other. The margin value is 0.8, which, through experiments, has been verified to achieve a good balance between intra-class aggregation and inter-class separation.

[0058] This application does not employ the traditional approach of image reconstruction followed by recognition. Instead, it directly performs feature representation and target recognition on the one-dimensional measurement sequence output by a near-infrared perovskite detector. Leveraging the high responsivity, low dark current, and high signal-to-noise ratio of perovskite detectors in the near-infrared band, the quality of the measurement sequence is improved. Combined with measurement domain feature learning methods, this achieves a synergistic enhancement of both device performance and recognition performance.

[0059] This application is described using Chinese character recognition as a specific example, but it is not limited to Chinese character recognition scenarios. It can also be applied to number recognition, industrial target detection, biometric recognition, and other near-infrared target classification tasks.

[0060] In this embodiment, the targets to be identified are multiple Chinese character categories. The selected perovskite detector uses tin-lead hybrid perovskite material, with a pixel size of 50μm, a measured dark current of 10⁻¹⁰ A / cm², a noise level of 10⁻⁶ A, and an external quantum efficiency ≥90%. The output wavelength range of the near-infrared broadband light source matches the peak response band of the perovskite detector. The digital micromirror device (DMD) has a resolution of 1024×768, and the number of encoded patterns generated by the Hadamard matrix used for slicing a cake is 8192. Several Chinese characters with different structural complexities are selected, and each character is used with different fonts, sizes, and slight rotations, generating a total of 200 samples. The sample set includes a training set and a validation set, and the sample set data includes similar sample pairs and dissimilar sample pairs. 150 samples are used as the training set to complete the pre-training of the Siamese network and the construction of the reference feature library, and to fix the network parameters; the other 50 samples are used as the validation set to complete the validation process and ensure the accuracy and confidence of target recognition.

[0061] In this embodiment, the emission band of the near-infrared broadband light source matches the response wavelength of the tin-lead perovskite detector. Thanks to the physical gain of the perovskite detector in the matched band, the shot noise limit of the detector is minimized. This allows the training of the Siamese network to achieve convergence with only a small number of samples.

[0062] The specific steps of Chinese character recognition in this embodiment include: first, placing the Chinese character pattern on the plane of the object and uniformly illuminating it using a near-infrared broadband light source; then, projecting the target-modulated light field onto the surface of the digital micromirror device, and using the digital micromirror device to load a digital coded aperture to perform spatial light modulation on the target.

[0063] The digital micromirror device is loaded with complementary codes generated by the Hadamard matrix for cutting a cake. Each group of Chinese character samples is modulated and sampled according to the same encoding order, and the encoding aperture number is transmitted to the data acquisition unit through a synchronous trigger signal.

[0064] The optical signal modulated by the digitally encoded aperture is received by a perovskite single-point detector. The detector converts the received optical signal into an electrical signal, and the corresponding measurement value is recorded by the data acquisition unit. In this embodiment, each Chinese character sample corresponds to a complete sequence of measurement values. The data processing and recognition module first establishes a response model based on the detector's dark current, responsivity, and noise parameters, and then corrects the original measurement values. The correction process includes dark current background subtraction, responsivity compensation, and normalization, further reducing the impact of detector drift and environmental noise on the experimental results.

[0065] The preprocessed measurement sequence is directly input into the trained feature extraction network, eliminating the need for image reconstruction. The feature extraction network compresses the original high-dimensional measurements into a 128-dimensional feature vector using a multi-layer encoder structure, and maps this feature vector to a pre-established feature space. The forward propagation process of the feature extraction network is as follows: the first input layer receives the measurement vector, compresses it to 1024 dimensions through a first fully connected layer, and then introduces a non-linear transformation using the ReLU activation function to enhance expressive power. The second layer further compresses the 1024-dimensional features to 256 dimensions, and again uses the ReLU activation function. The third layer finally outputs a 128-dimensional low-dimensional feature vector. The dimensionality reduction mapping relationship can be represented as h1+1=σ(W1h1+b1), where W1 is the weight matrix, b1 is the bias, and σ is the non-linear activation function. This encoding process, while preserving the core discriminative information of the target category, greatly removes redundant information from the measurement values, achieving efficient feature representation.

[0066] like Figure 3 After obtaining the 128-dimensional feature vector of the target, it is matched with a reference feature library pre-stored in the data processing unit. The distance is calculated using the Euclidean distance algorithm. The minimum distance *d* between the feature vector to be tested and all reference feature vectors in the target reference feature library is taken as the recognition result if it does not exceed a preset threshold. If the minimum distance exceeds the preset threshold, it is determined to be an unknown category target. The data processing and recognition module sends an unknown target signal to the rejection and expansion module and triggers the retraining-free expansion. After receiving the unknown target signal, the rejection and expansion module assigns a new category identifier to the unknown target, binds the low-dimensional feature vector of the unknown target to the category identifier, and stores it as a new reference feature vector in the target reference feature library, thus completing the triggered retraining-free expansion.

[0067] Figure 4 The confusion matrix results for the test dataset are given. In the confusion matrix, the horizontal axis represents the model's predicted class, the vertical axis represents the true class, the diagonal elements represent the proportion of correctly classified classes, and the off-diagonal elements represent the proportion of misclassified classes. Figure 4It can be seen that under 980nm near-infrared laser illumination, even though the imaging effect is affected by the environment, most samples are correctly classified, and the final recognition accuracy remains stable. The experimental results further demonstrate that the measurement sequence obtained after perovskite detector response parameter correction can more accurately reflect target feature information. Simultaneously, the feature space trained using a Siamese network has strong discriminative power, causing samples of the same category to cluster in the feature space, while samples of different categories maintain a large interval, thereby improving recognition accuracy and system robustness.

[0068] Furthermore, when new Chinese character categories need to be added, there is no need to retrain the entire neural network, which significantly reduces system maintenance costs and improves category expansion efficiency.

[0069] In summary, this embodiment verifies the feasibility of the non-reconstruction measurement domain target recognition method based on near-infrared perovskite detectors proposed in this invention. This method fully utilizes the advantages of perovskite detectors in the near-infrared band, such as high responsivity, high signal-to-noise ratio, and low-cost fabrication, to achieve direct mapping from measurement value sequences to target categories. While ensuring recognition accuracy, it avoids the additional computational overhead of image reconstruction, demonstrating good real-time performance, scalability, and promising engineering applications.

[0070] Secondly, this application provides a non-reconstruction measurement domain target recognition system based on a near-infrared perovskite detector, such as... Figure 2 As shown, the system is used to implement the non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector, including: The optical illumination and modulation module includes a near-infrared broadband light source, a beam expander / collimator lens group, an imaging lens, a dispersive element, a focusing lens, and a digital micromirror device, used to generate spectral-spatial hybrid coded patterns and perform spatial light intensity modulation; The photoelectric detection module includes a near-infrared perovskite single-point detector and a data acquisition unit, which is used to acquire modulated optical signals and generate a sequence of raw measurement values. The data processing and recognition module includes a data processing unit, which is used to preprocess the original measurement value sequence and reduce its dimension to a low-dimensional feature vector, calculate the minimum matching distance, send unknown target signals, and output target recognition results; The rejection and expansion module is used to receive and respond to unknown target signals, assign new category identifiers to unknown targets, and store the low-dimensional feature vectors of unknown targets into the target reference feature library.

[0071] Based on the above scheme, a near-infrared broadband light source with an output wavelength range of 780nm to 1100nm is used to emit a broadband near-infrared beam. The beam expander / collimating lens group, positioned after the near-infrared broadband light source, is used to expand and collimate the emitted beam to form a uniform illumination beam. The object plane, positioned after the beam expander / collimating lens group, is used to place the illuminated object, i.e., the target to be identified, allowing the collimated near-infrared beam to pass through the target and carry its spatial information. The imaging lens, positioned after the transmitting object, is used to converge the beam that has passed through the object to form an image. The dispersive element, positioned after the imaging lens, is used to spatially expand the imaging beam according to different wavelengths, forming a one-dimensional wavelength distribution for each wavelength component. The focusing lens, positioned after the dispersive element, is used to completely converge the dispersed beam to the surface of the digital micromirror device (DMD), so that different wavelength light components correspond to different spatial positions on the surface of the DMD, thereby forming a spectral-spatial hybrid coding pattern on the DMD surface. The system comprises a digital micromirror device (DMD) for focusing reflected light onto a perovskite single-point detector. The DMD has a surface bearing the spectral-spatial hybrid pattern, which is used to load a digitally coded aperture and modulate the spatial intensity of the hybrid pattern. Each time the coded aperture is refreshed, a trigger signal is sent to the data acquisition unit. The perovskite single-point detector is positioned in the reflected light path of the DMD, receiving the reflected light field modulated by the DMD and detecting its total intensity, converting the optical signal into an analog electrical signal. The perovskite single-point detector is made of near-infrared responsive perovskite material and has photoresponse capability in the 780nm–1100nm wavelength band. The data acquisition unit is connected to both the DMD and the perovskite single-point detector, receiving the trigger signal from the DMD and the analog signal output by the detector, and converting the analog signal into a digital measurement value sequence. The data processing unit is connected to the data acquisition unit, receiving the digital measurement value sequence and directly performing feature extraction and classification matching on the measurement value sequence in the measurement domain based on a pre-trained feature extraction model, outputting the recognition result without image reconstruction.

[0072] The data processing unit pre-stores the data storage program and a reference feature library trained using the same method.

[0073] Compared with existing technologies, this application improves the system signal quality under low light conditions by utilizing a near-infrared perovskite detector to achieve high-quality measurement value acquisition; it adopts a direct recognition strategy in the measurement domain, eliminating the need for image reconstruction, reducing computational complexity, and increasing recognition speed; through the collaborative design of the near-infrared perovskite detector and the non-reconstruction measurement domain recognition task, the measurement signal acquired by the detector can directly serve the target recognition process; it introduces a combination of a broadband near-infrared light source and a dispersive element, completing the joint encoding preprocessing of the spatial-spectral domain before the light signal reaches the DMD, significantly improving the dimensionality of the measurement value information; and it employs a recognition method that combines feature learning and feature matching, exhibiting good category expansion capabilities and system robustness.

[0074] Thirdly, this application provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector.

[0075] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as C, VHDL, Verilog, the object-oriented programming language Java, and the interpreted scripting language JavaScript.

[0076] This application is described with reference to flowchart illustrations of methods, apparatus, and computer program products according to embodiments of this application. It should be understood that each step in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0078] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 A process or multiple process diagrams specify the steps of a function.

[0079] In the description of this application, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0080] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0081] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector, characterized in that, include: A single-pixel optoelectronic hybrid neural network based on a near-infrared perovskite detector and having both optical and electrical end networks was constructed. Construct a target recognition sample set, wherein the sample set data includes similar sample pairs and dissimilar sample pairs; A single-pixel optoelectronic hybrid neural network was pre-trained using a target recognition sample set. After training, the network parameters were fixed and a target reference feature library was generated. By illuminating the target with optical light and modulating the optical path, a spectral-spatial hybrid code and a sequence of original measurement values ​​are formed. The original measurement value sequence is preprocessed based on the detector response model to obtain the standard measurement value sequence; The obtained standard measurement value sequence is input into a pre-trained single-pixel photoelectric hybrid neural network for feature extraction and target matching and recognition, and the target recognition result is output. The optical network is a coding layer used to achieve spectral-spatial hybrid coding; the electrical network is a decoding layer, including a preprocessing layer, a feature extraction layer, and a matching and recognition layer, used to identify targets from single-pixel detection signals. The optical illumination source is a near-infrared broadband light source; The target recognition process does not require the generation of a target reconstruction image throughout.

2. The non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector according to claim 1, characterized in that, The process of optically illuminating and modulating the optical path of the target to be identified to form a spectral-spatial hybrid code and a sequence of original measurement values ​​specifically includes: A broadband light beam is emitted using a near-infrared broadband light source. After passing through a lens group, the light beam converges to form an image after passing through the target to be identified. The transmitted beam carrying the target's spatial information is dispersed in space according to wavelength by a dispersive element to form a one-dimensional wavelength distribution. By using a focusing lens, all wavelength components after dispersion are converged onto the surface of the digital micromirror device to form a spectral-spatial hybrid coding pattern; A digital micromirror device is used to load a digital coded aperture to spatially modulate the intensity of a hybrid coded pattern and reflect it to a perovskite single-point detector. The received optical signal is converted into an electrical signal by a perovskite point detector, and the original measurement value sequence corresponding to the coded pattern is obtained.

3. The non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector according to claim 2, characterized in that, The method of using a digital micromirror device to load a digital coded aperture to spatially modulate the intensity of the hybrid coded pattern and reflect it to a perovskite single-point detector specifically includes: The digital micromirror device uses a complementary coding pattern generated by a Hadamard matrix for cutting a cake. It is divided into two groups, positive and negative, with the matrix values ​​of the two groups being opposite and complementary. The digital micromirror device is loaded with positive and negative complementary coded apertures in a predetermined order, and the near-infrared perovskite single-point detector sequentially collects the corresponding positive and negative measurement values. In this device, each row and column of pixels on the surface of the digital micromirror device corresponds to a specific wavelength range and spatial position. While loading the coded aperture, each micromirror unit synchronously modulates the spectral components of its corresponding spatial position independently, coupling the spatial position and wavelength information into a one-dimensional optical signal.

4. The non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector according to claim 2, characterized in that, The preprocessing of the original measurement value sequence based on the detector response model to obtain the standard measurement value sequence specifically includes: The data acquisition unit receives the analog electrical signal output by the perovskite point detector and converts it into a digital signal; The data processing algorithm corrects the acquired measurement sequence based on the physical response parameters of the perovskite detector, including dark current compensation, responsivity compensation, and normalization. The physical response parameters of the perovskite detector include dark current, noise level, quantum efficiency, responsivity, and near-infrared band response characteristics. Background subtraction is performed on the original measurement signal based on the detector's dark current parameters to obtain the measurement value sequence after eliminating background noise: M1 = M raw - I dark M raw For the original measurement sequence, I dark This is the dark current signal of the detector; Based on the measured response characteristics of the detector, a correction coefficient K is constructed to compensate for the measured values, resulting in the compensated measured value sequence: M2=M1 K; The standard measurement value sequence is obtained by normalizing the detector noise characteristics: M3=(M2-μ) / σ, where μ is the mean of the measurement signal and σ is the standard deviation of the measurement signal.

5. The non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector according to claim 4, characterized in that, The process of inputting the obtained standard measurement value sequence into a pre-trained single-pixel photoelectric hybrid neural network for feature extraction and target matching recognition, and outputting the target recognition result, specifically includes: The obtained standard measurement value sequence is input into a pre-trained single-pixel photoelectric hybrid neural network for feature extraction, so that the high-dimensional measurement value sequence is mapped into a low-dimensional feature vector. The low-dimensional feature vector is matched with the reference feature vector in the target reference feature library, the Euclidean distance is calculated, and the category with the smallest distance is output as the target recognition result. The Euclidean distance D is calculated using the Euclidean distance algorithm: ； in, represents the squared distance between the two input samples after being mapped by the neural network, and h represents the output vector of the hidden layer of the neural network. This represents the square of the L2 norm obtained by outputting the same network twice.

6. The non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector according to claim 5, characterized in that, The calculation of Euclidean distance, outputting the category with the smallest distance as the target recognition result, specifically includes: When the calculated minimum Euclidean distance is less than a preset threshold, the target recognition result corresponding to the minimum Euclidean distance is output; when the calculated minimum Euclidean distance is greater than or equal to the preset threshold, the target to be identified is determined to be an unknown target, an unknown target signal is sent to the system and retraining extension is triggered. After receiving an unknown target signal, the system assigns a new category identifier to the unknown target, binds the low-dimensional feature vector of the unknown target to the category identifier, and stores it as a new reference feature vector in the target reference feature library, thus completing the triggered retraining-free extension. During the retraining-free extension process, the network parameters remain frozen.

7. The non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector according to claim 1, characterized in that, The target identification sample set is constructed, and the sample set data includes similar sample pairs and dissimilar sample pairs; A single-pixel photoelectric hybrid neural network is pre-trained using a target recognition sample set. After training, the network parameters are fixed and a target reference feature library is generated. Specifically, this includes: The single-pixel optoelectronic hybrid neural network uses a twin network for feature extraction. The twin network consists of two encoders with identical structures and shared weight parameters. The two encoders receive two sets of measurement value sequences as input and output corresponding feature vectors. The pre-training process includes: The constructed target recognition sample set includes a training set and a validation set. The same type of sample pairs and different type of sample pairs in the training set are used as training data and input into a single-pixel photoelectric hybrid neural network as the target to be recognized. After optical illumination and optical path modulation, a spectral-spatial hybrid code is formed. By calculating the differential measurement values ​​and preprocessing the detector response model, a standard measurement value sequence is obtained and input into the twin network. In the Siamese network, the two encoders perform forward propagation on the two sets of input sequences through a multi-layer nonlinear mapping structure, reducing the original high-dimensional measurement values ​​to low-dimensional feature vectors. By continuously optimizing the network parameters through the comparative loss function until the loss function converges or reaches the preset maximum number of iterations, the network parameters are fixed, and the low-dimensional feature vectors of the samples are extracted using the network to construct the target reference feature library, thus completing the pre-training.

8. The non-reconstruction measurement domain target recognition method based on a near-infrared perovskite detector according to claim 7, characterized in that, The contrastive loss function is: L = y·D² + (1-y)·max(margin - D, 0)²; Where y is the label of the sample pair, with a value of 1 for similar sample pairs and a value of 0 for dissimilar sample pairs; D is the Euclidean distance between feature vectors; margin is a preset threshold parameter used to control the minimum interval of dissimilar sample pairs, with a value of 0.

8.

9. A non-reconstruction measurement domain target recognition system based on a near-infrared perovskite detector, used to implement the non-reconstruction measurement domain target recognition method according to any one of claims 1-8, characterized in that, include: The optical illumination and modulation module includes a near-infrared broadband light source, a beam expander / collimator lens group, an imaging lens, a dispersive element, a focusing lens, and a digital micromirror device, used to generate spectral-spatial hybrid coded patterns and perform spatial light intensity modulation; The photoelectric detection module includes a near-infrared perovskite single-point detector and a data acquisition unit, which is used to acquire modulated optical signals and generate a sequence of raw measurement values. The data processing and recognition module includes a data processing unit, which is used to preprocess the original measurement value sequence and reduce its dimension to a low-dimensional feature vector, calculate the minimum matching distance, send unknown target signals, and output target recognition results; The rejection and expansion module is used to receive and respond to unknown target signals, assign new category identifiers to unknown targets, and store the low-dimensional feature vectors of unknown targets into the target reference feature library.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the non-reconstruction measurement domain target recognition method as described in any one of claims 1-8.

Images