Single-pixel target recognition method based on optoelectronic hybrid neural network

Abstract

The application provides a fast target recognition method based on an optoelectronic hybrid neural network, and solves the problem of low information acquisition efficiency of a target recognition system, and comprises the following steps: an optical network is designed to realize cyclic coding single-pixel detection, and an electrical network is simultaneously constructed to decode the single-pixel detection signal and acquire object information; a data set used for training the hybrid neural network is selected according to an application scene; the optoelectronic hybrid neural network is trained by using the acquired data set; a cyclic coding mask obtained by training is processed, and a single-pixel detection system based on the coding mode is built; object information is collected by a single-pixel detector after being modulated by the cyclic coding mask; and finally, the pre-trained electrical neural network is used to recognize the object information from the detection signal. The application cooperatively uses the cyclic coding single-pixel detection method and the artificial intelligence technology, and can realize accurate prediction of the category of a high-speed moving object to be detected in a special waveband and a weak light environment.

Description

Technical Field

[0001] This invention relates to the fields of optical information processing and artificial intelligence, and in particular to a single-pixel image recognition method based on a hybrid optoelectronic neural network. Background Technology

[0002] Point-to-point imaging technology projects target object information onto an image sensor through a lens, and the image is obtained through photoelectric conversion. Target recognition can then be achieved by employing intelligent algorithms such as deep learning. Although this recognition method has been widely used in many recognition tasks, it relies on traditional imaging techniques to pre-acquire scene image information, thus limiting the system's recognition capability to the performance of the image sensor. Traditional area array photodetectors suffer from limitations in photoelectric conversion speed, spectral response range, and sensitivity, thereby restricting the application of existing target recognition technologies in typical scenarios such as high-speed motion scenes, special wavelengths, and low-light environments.

[0003] All-optical neural networks can directly process optical signals, avoiding the photoelectric conversion process and overcoming the performance limitations of image sensors, enabling light-speed recognition of target scenes. However, existing optical neural networks suffer from difficulties in nonlinear transformation, limited parameter quantity, poor reconfigurability, and high development costs, which greatly restrict the practical application of this method.

[0004] Compared to traditional target recognition technologies that acquire the entire image and all-optical neural networks that don't acquire any image at all, a compromise is to only acquire information useful for target recognition. In fact, the information needed for target recognition is often only a part of the complete image information of the scene. We can record only the scene information needed for image recognition, thereby improving the information acquisition efficiency of the perception system.

[0005] Single-pixel cameras, by leveraging the sparsity of scenes, can effectively avoid acquiring redundant information and have demonstrated superior performance in compressed imaging and image-free recognition. Furthermore, single-pixel detectors also offer advantages in photoelectric conversion speed, spectral response range, and sensitivity. Therefore, target recognition methods based on single-pixel detectors hold promise for solving target recognition problems in typical scenarios such as high-speed motion, special wavelengths, and low-light environments. However, this approach requires recovering object information from light field intensity fluctuation sequences obtained after multiple encodings of the illumination scene. Therefore, the speed of information acquisition is limited by the number of samples required for target recognition and the refresh rate of the light field modulation device. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a single-pixel target recognition method and device based on a photoelectric hybrid neural network. This method does not require acquiring complete image information of the object, and requires less data acquisition and shorter acquisition time, making it suitable for recognizing high-speed moving objects.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A single-pixel target recognition method based on a photoelectric hybrid neural network is characterized by the following steps:

[0009] S1: Construct a single-pixel optoelectronic hybrid neural network;

[0010] S2: Select the training dataset;

[0011] S3: Train the optoelectronic hybrid neural network using the training dataset;

[0012] S4: Process the encoding device based on the model parameters obtained from training;

[0013] S5: Build a single-pixel detection system based on this encoding device;

[0014] S6: Use the constructed single-pixel detection system to obtain the encoded object information;

[0015] S7: Use a pre-trained electronic neural network to identify object categories from single-pixel detection signals;

[0016] S8: For a new scene to be identified, repeat steps S6 to S7.

[0017] Furthermore, in one embodiment of the present invention, the single-pixel optoelectronic hybrid neural network described in step S1 is a self-encoding architecture, wherein the optical end network is the encoding layer, the electrical end network is the decoding layer, and the bottleneck layer is the output of the optical end network, which is also the input of the electrical end. The optical end network is used to design the encoding mask of the single-pixel detection system, and the electrical end network is used to realize target recognition from the single-pixel detection signal.

[0018] Furthermore, in one embodiment of the present invention, the optical end network in step S1 is an encoding process for a single-pixel detection system based on a cyclic mask. The structure of the cyclic mask, i.e., the encoding matrix of the single-pixel detection system, can be adaptively optimized during training, and its dimension is determined by the image size and sampling rate.

[0019] Furthermore, in one embodiment of the present invention, the optical end network in step S1 is implemented by circular convolution, and the convolution kernel is the structure of the circular mask to be learned.

[0020] Furthermore, in one embodiment of the present invention, the convolution kernel is in a binary form of 0 / 1, which facilitates the physical implementation of the optical end network.

[0021] Furthermore, in one embodiment of the present invention, the electrical terminal network in step S1 identifies the object category based on the single-pixel detection signal of the target, including a preprocessing layer, a physically guided decoding layer, and an image recognition module containing a multi-layer neural network structure.

[0022] Furthermore, in one embodiment of the present invention, the preprocessing layer standardizes the single-pixel detection signal, enabling the subsequent recognition network to focus on the fluctuations of the single-pixel detection signal, thus making it suitable for target recognition under different illumination levels.

[0023] Furthermore, in one embodiment of the present invention, the physically guided decoding layer uses a cyclic coding mask in the optical end network to perform intensity differential correlation operation with the standardized single-pixel detection signal. The operation result can be regarded as the extraction of scene image features. The process uses physical methods including a cyclic coding mask (forward physical model) and intensity differential correlation operation rules (image feature reconstruction method). This process converts the one-dimensional detection signal into two-dimensional image features.

[0024] Furthermore, in one embodiment of the present invention, the image recognition module calculates and outputs the object category based on the image features extracted by the physically guided decoding layer.

[0025] Furthermore, in one embodiment of the present invention, the training data in step S2 includes object images and their categories, and the selection of the training dataset depends on the requirements of the recognition task.

[0026] Furthermore, in one embodiment of the present invention, the training process of the optoelectronic hybrid neural network in step S3 takes the object image in the selected dataset as input, the category as label, the image is processed by the optical end network to obtain a one-dimensional light intensity fluctuation sequence, and then outputs the prediction result through the electrical end network. The error between the prediction result and the label data is calculated using cross-entropy as the loss function. The training process is to adjust the parameters of the optical end and the electrical end network to minimize the prediction error.

[0027] Furthermore, in one embodiment of the present invention, the fabrication of the encoding device in step S4 is intended to create a modulation device to enable the recurrent mask obtained from the optical end network training to quickly encode the target object.

[0028] Furthermore, in one embodiment of the present invention, the encoding device is a rotating disk with a rigid transparent material as the substrate and a film coated on the circular area of ​​the substrate. The coating structure is a training-derived cyclic mask structure. The encoding process based on single-pixel detection can be realized by rotating the disk with a motor.

[0029] Furthermore, in one embodiment of the present invention, a cyclic mask structure trained under different conditions (such as different datasets) can be coated at different radii of the rotating disk to obtain multiple concentric rings. By selecting different rings to illuminate objects, the same rotating disk can be used to encode target objects under different conditions.

[0030] Furthermore, in one embodiment of the present invention, the process of obtaining the cyclic mask structure coating is as follows: coating on a transparent substrate, and then using etching technology to create a structured pattern on the film.

[0031] Furthermore, in one embodiment of the present invention, the cyclic mask is arranged in multiple integer periods to fill the entire ring. The cyclic mask is implemented by circular convolution, so masks of different periods on the ring can be connected end to end. The motor drives the mask to rotate one mask cell to achieve one refresh, and rotating one cycle can achieve one recording of the signal required for target recognition.

[0032] Furthermore, in one embodiment of the present invention, the single-pixel detection system in step S6 includes a power supply, a light source, a data acquisition card, a single-pixel detector, a computer, a rotating disk, a motor, a lens, an aperture, and other devices. The core device is a rotating disk used to perform cyclic encoding on the target object.

[0033] Furthermore, in one embodiment of the present invention, the single-pixel detection system described in step S6 can image transmissive targets and reflective targets, and the object encoding can be performed on the object plane (structured illumination) or on the image plane (structured detection).

[0034] Furthermore, in one embodiment of the present invention, the single-pixel detection system in step S6 acquires target object information by synchronizing the motor and the single-pixel detector. The position of the motor rotation determines the pattern of the currently encoded object information. The single-pixel detector needs to record the corresponding light intensity signal at the corresponding pattern. The recorded light intensity signal is the output result of the optical neural network.

[0035] Furthermore, in one embodiment of the present invention, the pre-trained electronic neural network used in step S7 is the one trained in step S3.

[0036] Furthermore, in one embodiment of the present invention, step S7 outputs a prediction result of the scene category to be tested based on the single-pixel detection result.

[0037] A single-pixel target recognition device based on a photoelectric hybrid neural network, characterized in that it includes an illumination module, a control module, a modulation module, a detection module, and a calculation module;

[0038] The lighting module includes a lighting source and a beam shaping device for illuminating the scene under test.

[0039] The control device is connected to the modulation module and the detection module to control the synchronization of the rotating motor and the single-pixel detector.

[0040] The modulation module includes a beam shaping device, a rotating disk, and a motor. The rotating disk uses a transparent rigid material as a substrate, and a membrane with a trained cyclic mask structure is deposited on the circular ring of the substrate. The motor drives the rotating disk to refresh the coded mask for cyclic encoding of single-pixel detection every time it rotates through one mask pixel size. An adjustable adapter is used between the rotating disk and the motor to ensure that the center of the ring is aligned with the rotation center of the motor and that the disk plane is perpendicular to the motor shaft.

[0041] The detection module includes a beam collection system, a single-pixel detector, and a data acquisition card, used to acquire modulated light intensity fluctuation signals.

[0042] The computing module includes units for computing, storage, and display, and is connected to the detection module to predict the category of the target scene from the light intensity fluctuation signal obtained by the detection module.

[0043] Each module selects one of four single-pixel detection system devices based on the type of scene to be detected and the modulation method: transmissive object plane modulation, transmissive image plane modulation, reflective object plane modulation, and reflective image plane modulation.

[0044] The refresh rate S of the modulation module of the single-pixel detection system is jointly determined by the diameter D of the circular mask, the motor rotation speed T, and the pixel size P of the mask structure, and the calculation formula is as follows:

[0045]

[0046] The target recognition rate of the single-pixel detection system is determined by the refresh rate S and the number of samples. The ratio is determined by this.

[0047] The coded mask structure is used to encode object information through beam shaping devices in the illumination module, modulation module, and detection module, and the total light intensity information after encoding is recorded by a photodetector.

[0048] Compared with existing technologies, the advantages of this invention are as follows: This invention combines single-pixel detection technology with artificial intelligence technology by designing a hybrid optoelectronic neural network. It uses a rotating disk to achieve efficient encoding of object information by the optical network, and then uses an electrical network to achieve high-accuracy target recognition from the single-pixel detection signal. Compared with traditional target recognition technology based on two-dimensional images (pure electronic neural networks), this invention avoids recording target object images, reduces the recording of redundant information, and improves information acquisition efficiency, making it suitable for recognizing high-speed moving objects. Compared with target recognition technology based on all-optical neural networks, this invention combines the advantages of both optical and electrical networks, achieving a target recognition system with high accuracy, strong stability, and low cost. Furthermore, due to the advantages of the modulation device and single-pixel detection, this invention has broad application prospects in target recognition problems in special wavelength bands and low-light scenarios. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the optoelectronic hybrid neural network structure provided in the embodiments of this application.

[0050] Figure 2 This is a schematic diagram of the method for implementing cyclic circular convolution provided in the embodiments of this application.

[0051] Figure 3 This is a schematic diagram of the method for preparing a rotating disk provided in the embodiments of this application.

[0052] Figure 4 This is a schematic diagram of a single-pixel detection system provided in an embodiment of this application.

[0053] Figure 5 This is a schematic diagram of the target recognition system provided in the embodiments of this application.

[0054] Figure 6 This is a flowchart of a single-pixel image recognition method based on a photoelectric hybrid neural network provided in the embodiments of this application.

[0055] Among them, 1-Illumination source; 2-Beam shaping device; 3-Rotating disk; 4-Microscopic objective; 5-Bullet lens; 6-Target object; 7-Light collecting lens; 8-Barrel detector; 9-Power supply; 10-Motor; 11-Data acquisition card; 12-Computer. Detailed Implementation

[0056] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments, but this should not be construed as limiting the scope of protection of the present invention.

[0057] The target recognition method based on optoelectronic hybrid neural networks specifically includes the following steps:

[0058] Step S1: Construct a single-pixel photoelectric hybrid neural network. For example... Figure 1 As shown, the optical end of the optoelectronic hybrid neural network implements the cyclically encoded single-pixel detection process, while the electrical end distinguishes the object category from the intensity information of the single-pixel detection. Specifically, let the input data be... The optical end network is R ω in The kernel is a circular convolution (β is the selected sampling rate, i.e., the ratio of the number of samples to the total number of pixels in the image), x is processed by R ω Light intensity fluctuations I are obtained by single-layer circular convolution processing

[0059]

[0060] Here, circular convolution is indirectly implemented through linear convolution, and the specific implementation process is as follows: Figure 2 As shown, the input image x is first translated and padded with zeros to obtain the input to the convolutional layer. Then convolve with ω to obtain This is the output of the optical neural network. To facilitate subsequent implementation of circular convolution using a rotating disk, such as... Figure 2 As shown, copy the first N columns of ω and concatenate them at the end, then delete the first column of ω to obtain the complete encoding mask. By ω f This can be obtained by taking every pixel from left to right. A coding mask of size M×N is used. Considering that the coding mask cannot fill an entire circular ring, such as... Figure 2 As shown, by copying the encoding mask and connecting the first and last parts together, the space of the rotating disk can be fully utilized to achieve cyclic encoding.

[0061] For electrical network structures, such as Figure 1 As shown, the electrical terminal neural network takes I as input, and after processing through a normalization layer, a physical guidance decoding layer, and an image recognition network layer, it outputs the recognition result.

[0062] The normalization layer processes I as follows:

[0063]

[0064] in Let σ be the mean of I, and σ be the standard deviation of I.

[0065] The physical guidance decoding layer uses ω f and The differential correlation operation is performed, and the specific calculation formula is as follows:

[0066]

[0067] in ω i From ωf Extracting from the middle, moving one pixel from left to right, taking an M×N encoding mask, and then spacing them i pixels apart to obtain... A total of A mask.

[0068] The image recognition network layer can be selected from one of the classification networks such as VGG, ResNet, DenseNet, Inception, MobileNet, and EfficientNet.

[0069] Step S2: Select the training dataset. Choose a dataset for object recognition based on the task at hand. This can be from various open-source datasets such as MNIST, Fashion-MNIST, ImageNet, CelebA, etc., or from a user-created dataset. Each set of data in the dataset includes scene images and their corresponding labeled categories.

[0070] Step S3: Train the optoelectronic hybrid neural network using the training dataset. Divide the dataset selected in Step S2 into training, validation, and test sets in appropriate proportions. Using the training dataset, adjust the optical and electrical network parameters using stochastic gradient descent to minimize the error between the optoelectronic hybrid neural network output and the labeled data. Training consists of multiple epochs. The performance of the current model on the validation set is used to evaluate whether training needs to be terminated to prevent overfitting.

[0071] Step S4: Fabricate the encoding device based on the trained model parameters. Based on the training results obtained in Step S3, extract the cyclic encoding mask ω from the optical network. f It is then placed on a rotating disk. The rotating disk uses a transparent rigid material as a substrate, and a film is coated on a circular area on the substrate. The coating structure is a training-derived cyclic mask structure ω. f .

[0072] The coating process requires a rectangular matrix ω f Generate ring structures, such as Figure 3 As shown, ω f Each column corresponds to a different rotation angle, ω f Each row corresponds to a different radius, and then ω is recalculated on the annulus. f All structures are arranged in a circular pattern. The circular mask is arranged in multiple integer periods to fill the entire ring. The circular mask is implemented by circular convolution, so masks of different periods on the ring can be connected end to end. The motor drives the mask to rotate one mask cell to achieve one refresh, and one rotation cycle can achieve one recording of the signal required for target recognition.

[0073] The refresh rate S of the encoding device is determined by the ring diameter D, the motor speed T, and the mask pixel size P, using the following formula:

[0074]

[0075] For example, if the diameter of the ring is D = 200mm, the motor speed is T = 20r / s, the mask pixel size is P = 4μm, and the refresh rate is S≈3.14MHz.

[0076] The motor is connected to the rotating disk via an adjustable adapter, which is used to adjust the alignment of the disk center with the motor rotation center.

[0077] Step S5: Construct a single-pixel detection system based on this encoding device. For example... Figure 4 As shown, the single-pixel detection system includes: 1-illumination source; 2-beam shaping device; 3-rotating disk; 4-microscope objective; 5-tunnel lens; 6-target object; 7-light-collecting lens; 8-barrel detector; 9-power supply; 10-motor; 11-data acquisition card; 12-computer, etc. The cooperating structure includes four forms: transmissive object plane modulation, transmissive image plane modulation, reflective object plane modulation, and reflective image plane modulation.

[0078] Step S6: Obtain the encoded object information using the constructed single-pixel detection system. The object is placed... Figure 4 At the target object (6) shown, by turning on the light source and motor, the cyclic mask pattern on the rotating disk is used to encode the target object information, and the intensity of the encoded signal is recorded sequentially using a single pixel detector. The motor rotates past one cyclic mask pixel and records one light intensity value. According to the set sampling rate, the motor rotates past... Each pixel yields a set of light intensity fluctuation signals I used for target recognition. The signal acquisition and recognition speed of the detection system is:

[0079]

[0080] For example, using a rotating disk with a refresh rate of S≈3.14MHz, for a 100×100 image, with a sampling rate of 0.1, the single-pixel detection system can acquire the light intensity fluctuation signal I at a rate of 3.14kHz. For some sparse target objects, the sampling rate can be as low as 0.01, and the rate at which the single-pixel detection system acquires the light intensity fluctuation signal I can be further increased to 31.4kHz.

[0081] Step S7: Use a pre-trained electronic neural network to identify the object category from the single-pixel detection signal. Based on the light intensity fluctuation signal I obtained in step S6, input it into the pre-trained image recognition module at the electronic end. After processing through the normalization layer, the physical guided decoding layer, and the image recognition network layer, the prediction result of the scene category is obtained.

[0082] Step S8: For a new scene to be identified, repeat steps S6 to S7. For example... Figure 5 As shown, the scene to be identified is collected by a single-pixel detection system to obtain light intensity fluctuations, and the light intensity fluctuations are input into the image recognition module to obtain the prediction result of the scene category.

[0083] The foregoing explanation of the target recognition method based on photoelectric hybrid neural network also applies to the target recognition device based on photoelectric hybrid neural network, and will not be repeated here.

[0084] According to a preferred embodiment of the present invention, a target recognition method and apparatus based on a hybrid photoelectric neural network are proposed. The optical end uses a single-pixel detection method based on circular convolution, while the electrical end uses image recognition technology to predict object categories from light intensity fluctuations. This method does not require pre-acquiring images, but only needs to acquire a small amount of information related to image recognition, facilitating information acquisition, storage, and transmission. The circular convolution at the optical end is implemented through a rotating disk, which has an ultra-high refresh rate and a wide operating wavelength, enabling rapid acquisition of light intensity signals for target recognition. It can be used for target recognition under high-speed motion, special wavelengths, and low-light conditions.

[0085] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, several equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or purpose, should be considered within the scope of protection of the present invention.

Claims

1. A single-pixel target recognition method based on an optoelectronic hybrid neural network, characterized in that, The method comprises the following steps: S1.1: constructing a single-pixel optoelectronic hybrid neural network with an optical end network and an electrical end network; S1.2: establishing a single-pixel target recognition data set, which comprises object images and categories corresponding to the object images; S1.3: training the single-pixel optoelectronic hybrid neural network by using the single-pixel target recognition data set; S1.4: processing an encoder device according to a cyclic mask obtained by training, so as to realize encoding of the object images; S1.5: constructing a single-pixel detection system based on the encoder device, which is used to realize cyclic convolution of the object images by the optical end network, so as to obtain an intensity fluctuation signal; S1.6: obtaining encoded object information by using the single-pixel detection system; S1.7: outputting a category of a to-be-detected object according to the object images in the single-pixel detection signal by using the single-pixel optoelectronic hybrid neural network trained in step S3; S1.8: repeating steps S6 to S7 for a new to-be-recognized scene; The single-pixel optoelectronic hybrid neural network is a self-encoding architecture, the optical end network is an encoding layer, the electrical end network is a decoding layer, and a bottleneck layer is an output of the optical end network and an input of the electrical end network; The optical end network is a single-pixel detection technology based on a cyclic mask, takes an object image as an input, and takes an intensity fluctuation signal of light intensity as an output, and is used to design an encoding mask of a single-pixel detection system; the electrical end network is used to predict a category of an object from the intensity fluctuation signal of light intensity output by the optical end network, and comprises a preprocessing layer, a physically guided decoding layer, and an image recognition network layer, and is used to realize target recognition from the single-pixel detection signal.

2. The single-pixel target recognition method of claim 1, wherein, The cyclic mask, i.e., an encoding matrix of the single-pixel detection system, can be adaptively optimized in a training process, and a dimension thereof is determined by an image size and a sampling rate.

3. The single-pixel target recognition method of claim 1, wherein, The preprocessing layer is used to standardize the single-pixel detection signal, so that a subsequent recognition network pays attention to fluctuation of the single-pixel detection signal, and the single-pixel detection signal is suitable for target recognition under different light levels; the physically guided decoding layer is used to perform intensity differential correlation operation on the cyclic encoding mask in the optical end network and the standardized single-pixel detection signal, and an operation result can be regarded as extraction of image features of a scene, and a process uses physics including the cyclic encoding mask (a forward physical model) and an intensity differential correlation operation rule (an image feature reconstruction method), and the process converts a one-dimensional detection signal into two-dimensional image features; the image recognition network layer is used to calculate and output a category of an object according to the image features extracted by the physically guided decoding layer.

4. The single-pixel target recognition method of claim 1, wherein, In step S3, the single-pixel optoelectronic hybrid neural network is trained by using the single-pixel target recognition data set, specifically as follows: S6.1: dividing the single-pixel target recognition data set into a training set, a verification set and a test set; each group of data in the data set comprises an image and a category corresponding to the image; S6.2: using the training set, selecting part of data pairs in the training set each time, taking pictures in the data as inputs of the optoelectronic hybrid neural network, and taking corresponding categories as labels corresponding to the inputs; S6.3 The input image is processed by the optoelectronic hybrid neural network to obtain a predicted category, and the predicted category is compared with the true label to obtain a loss function; the gradients of the learnable parameters of the optoelectronic hybrid neural network are calculated according to the back propagation algorithm, and then the parameters are updated in the gradient descent manner; a binary constraint is introduced to the parameters of the optical end network, and the update of the parameters can only be selected between 0 and 1; S6.4 After each training round, the actual performance of the current network is evaluated using the validation set, and the model with the best performance on the validation set is selected as the final training result; The performance on the test set is used as a reference for the actual performance of the network model; S6.5 When the performance of the obtained model on the test set meets the requirements, the training process is completed; if the requirements are not met (the model has a low recognition accuracy for the test set images), the network hyperparameters are further adjusted, and then the training is performed again, including the given sampling rate, learning rate, and network structure; S6.6 Repeat the above process, and when the performance of the trained model on the test set meets the requirements, the current trained network model is considered as the final training result, and the training is completed.

5. The single-pixel target recognition method of claim 4, wherein, In the training process, the learnable parameters of the optical end network are binarized to realize binary circular mask coding; the trained optical end network obtains a circular mask, which is implemented on a rotating disc in hardware, and a plurality of circular masks are connected end to end to fill a circular ring on the disc, and circular rings with different radii are used to realize the circular masks obtained under different conditions.

6. The single-pixel target recognition method of claim 1, wherein, The encoder is a rotating disc with a rigid transparent material as the base, and a film is coated on the circular ring range on the base, and the film structure is the circular mask structure obtained by training. The rotation of the disc driven by the motor can realize the coding process of single-pixel detection based on the circular mask.

7. The single-pixel target recognition method of claim 6, wherein, At different radii of the rotating disc, the circular mask structures obtained under different conditions are coated to obtain a plurality of concentric circular rings. By selecting different circular ring illuminations, the same rotating disc can be used to realize the coding of target objects under different conditions. The circular mask is arranged by a plurality of integer periods to fill the entire circular ring. The circular mask is implemented by circular convolution, so that the different period masks on the circular ring can be connected end to end. The motor drives the mask to rotate one mask pixel to realize one refresh, and rotates one period to realize the recording of the signal required for target recognition once.

8. The single-pixel target recognition method of claim 1, wherein, The single-pixel detection system includes a power supply, a light source, a data acquisition card, a single-pixel detector, a computer, an encoder, a motor, a lens, a support structure, and a diaphragm. The encoder is a rotating disc used for circular coding of target objects. The power supply is connected to the computer, the light source, the data acquisition card, and the motor for power supply. The data acquisition card is connected to the single-pixel detector and the motor for synchronous acquisition under the control of the computer program. The encoder is connected to the motor. The lens and the diaphragm are placed in the light path to shape the light beam. For the object plane coding scheme, the light source is modulated by the encoder to form a structured light to illuminate the object. For the image plane coding scheme, the echo after the object is illuminated by the light source is modulated by the encoder. The support structure fixes each component.

9. The single-pixel target recognition method of claim 5, wherein, The single-pixel detection system can image the transmissive target and the reflective target, and the encoding of the object can be on the object plane (structured illumination) or on the image plane (structured detection); the single-pixel detection system acquires the target object information in dependence on the synchronization of the motor and the single-pixel detector, the position of the motor rotation determines the current encoding object information pattern, and the single-pixel detector needs to record the corresponding light intensity signal at the corresponding pattern, and the recorded light intensity signal is the output result of the optical end neural network.

10. The single-pixel target recognition method of claim 1, wherein, The single-pixel detection system in the step S1.5 comprises: an illumination module for illuminating the scene to be measured; a control module for controlling the synchronization of the rotating motor and the single-pixel detector; a modulation module connected to the control module, comprising a rotating disc and a motor; the rotating disc is based on a transparent rigid material, and a trained cyclic mask is plated on the annular base of the rotating disc; the motor drives the rotating disc to rotate by one pixel size of the mask to realize the cyclic encoding single-pixel detection encoding mask refresh once; an adjustable adapter is used between the rotating disc and the motor to realize the alignment of the center of the annular ring and the rotating center of the motor and the perpendicularity of the disc plane and the motor rotating shaft; a detection module comprising a light beam collection system, a single-pixel detector and a data acquisition card, for collecting and obtaining the modulated light intensity fluctuation signal; a calculation module connected to the detection module, for predicting the category of the target scene to be measured from the light intensity fluctuation signal obtained from the detection module.

11. The single-pixel target recognition method of claim 10, wherein, The refresh rate of the modulation module satisfy the following conditions: wherein, is the circular ring mask diameter, is the motor rotation speed, is the mask structure pixel size; The target recognition rate of the single-pixel detection system is determined by the ratio of the refresh rate and the number of samples .

Images