Single-pixel target identification method based on photoelectric hybrid neural network

Abstract

The invention provides a fast target identification method based on a photoelectric hybrid neural network, which solves the problem of low information acquisition efficiency of a target identification system, and comprises the following steps: designing an optical network to realize cyclic coding single-pixel detection, and constructing an electrical network to decode a single-pixel detection signal to acquire information of an object to be detected; selecting a data set for training the hybrid neural network according to an application scene; training the photoelectric hybrid neural network by using the acquired data set; processing and training the obtained cyclic coding mask and constructing a single-pixel detection system based on the coding mode; the object information is modulated by the cyclic coding mask and then is collected by the single-pixel detector; and finally, identifying object information from the detection signal by using a pre-trained electronic neural network. According to the method, a single-pixel detection method of loop coding and an artificial intelligence technology are cooperatively used, and accurate prediction of the type of the to-be-detected object moving at a high speed can be realized in a special wave band and weak light environment.

Description

Technical Field

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

[0002] Point-to-point imaging technology is used to project the target object information onto the image sensor through a lens, and the object image is obtained through photoelectric conversion. Target recognition can be achieved with the help of intelligent algorithms such as deep learning. Although this recognition method has been widely used in many recognition tasks, it relies on traditional imaging technology to obtain scene image information in advance, so the system recognition ability is limited by the performance of the image sensor. Traditional array photoelectric detectors have problems such as photoelectric conversion speed, spectral response range, and limited sensitivity, which in turn limits the application of existing target recognition technology in typical scenes such as high-speed motion scenes, special bands, and low-light environments.

[0003] All-optical neural networks can directly process light signals, avoiding the photoelectric conversion process, breaking through the performance limitations of image sensors, and can achieve light-speed recognition of target scenes. However, existing optical neural networks have problems such as difficulty in nonlinear transformation, limited parameters, poor reconfigurability, and high development costs, which greatly limit the practical application of this method.

[0004] Compared with the traditional target recognition technology that collects complete images and the all-optical neural network that does not collect images at all, a compromise solution is to only collect information that is useful for target recognition. In fact, the information required for target recognition is often only a part of the complete image information of the scene. We can only record the scene information required for image recognition, thereby improving the information acquisition efficiency of the perception system.

[0005] Single-pixel cameras can effectively avoid the acquisition of redundant information by utilizing the sparsity of the scene, and have demonstrated superior performance in compressed imaging and non-imaging recognition. In addition, single-pixel detectors also have advantages in photoelectric conversion speed, spectral response range, and sensitivity. Therefore, the target recognition method based on single-pixel detectors is expected to solve the problem of target recognition in typical scenes such as high-speed motion, special bands, and low-light environments. However, this scheme needs to recover object information from the light field intensity fluctuation sequence obtained after multiple encoding illumination scenes, so the speed of acquiring information 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] In order to solve the above technical problems, the present invention proposes a single-pixel target recognition method and device based on an optoelectronic hybrid neural network. This method does not need to obtain the complete image information of the object, and the amount of data required to be collected is small and the collection time is short. It can be used for the recognition of high-speed moving objects.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A single-pixel target recognition method based on an optoelectronic hybrid neural network is characterized in that it includes the following steps:

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

[0010] S2: Select training data set;

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

[0012] S4: Processing the encoding device according to the model parameters obtained through training;

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

[0014] S6: Obtain the encoded object information using the constructed single-pixel detection system;

[0015] S7: Object category recognition from single-pixel detection signals using pre-trained electronics neural networks;

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

[0017] Furthermore, in one embodiment of the present invention, the single-pixel optoelectronic hybrid neural network in step S1 is a self-encoding architecture, in which the optical end network is the encoding layer, the electrical end network is the decoding layer, and the bottleneck layer is the optical end network output, that is, the electrical end input. The optical end network is used to realize the design of 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 a single-pixel detection system encoding process based on a cyclic mask, and the structure of the cyclic mask, i.e., the encoding matrix of the single-pixel detection system, can be adaptively optimized during the training process, and its dimension is determined by the image size and the 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 cyclic 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 is convenient for the physical implementation of the optical end network.

[0021] Furthermore, in one embodiment of the present invention, the electrical end 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 including a multi-layer neural network structure.

[0022] Furthermore, in one embodiment of the present invention, the preprocessing layer normalizes the single-pixel detection signal so that the subsequent recognition network focuses on the fluctuation of the single-pixel detection signal, 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 the cyclic coding mask in the optical end network to perform intensity differential association operation with the standardized single-pixel detection signal. The operation result can be regarded as the extraction of scene image features. The physics used in the process includes cyclic coding mask (forward physical model) and intensity differential association operation rules (image feature reconstruction method). This process converts the one-dimensional detection signal into a two-dimensional image feature.

[0024] Furthermore, in one embodiment of the present invention, the image recognition module calculates and outputs object categories based on 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 data set 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 data set as input and the category as a label. The image is processed by the optical end network to obtain a one-dimensional light intensity fluctuation sequence, and then output by the electrical end network to obtain a prediction result. The cross entropy is used as the loss function to calculate the error between the prediction result and the label data. The training process adjusts the optical and electrical network parameters to minimize the prediction error.

[0027] Furthermore, in one embodiment of the present invention, the processing of the encoding device in step S4 is intended to produce a modulation device to implement rapid encoding of the target object using a cyclic mask obtained by optical end network training.

[0028] Furthermore, in one embodiment of the present invention, the encoding device is a rotating disk, the disk has a rigid transparent material as a substrate, and is coated in a circular range on the substrate, the coating structure is a trained circular mask structure, and the rotation of the disk driven by a motor can realize the encoding process of single-pixel detection based on the circular mask.

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

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

[0031] Furthermore, in one embodiment of the present invention, the cyclic mask is arranged through multiple integer cycles to fill the entire circle, and the cyclic mask is implemented through circular convolution, so that masks of different cycles on the circle can be connected end to end. The motor drives the mask to rotate one mask pixel to achieve a refresh, and rotates one cycle to achieve a record of the signal required for target identification.

[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, and the core component is a rotating disk for realizing cyclic encoding of the target object.

[0033] Furthermore, in one embodiment of the present invention, the single pixel detection system in step S6 can image transmissive targets and reflective targets, and the encoding of objects 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 obtains the target object information by relying on the synchronization of the motor and the single-pixel detector. The rotation position of the motor determines the pattern of the current 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 end neural network.

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

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

[0037] A single-pixel target recognition device based on an optoelectronic 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, and is used to illuminate the scene to be measured.

[0039] The control device is connected with 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, a motor, etc. The rotating disk is based on a transparent rigid material, and a circle of a film having a trained cyclic mask structure is plated on the circular ring of the substrate. The motor drives the rotating disk to refresh the coding mask for cyclic coding single pixel detection every time it rotates through a mask pixel size. An adjustable adapter is used between the rotating disk and the motor to align the center of the ring with the rotation center of the motor and to make the plane of the disk perpendicular to the motor shaft.

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

[0042] The calculation module includes calculation, storage, display and other units, is connected to the detection module, and predicts the category of the target scene to be measured from the light intensity fluctuation signal obtained by the detection module.

[0043] Each module selects any one of four single-pixel detection system devices, namely, transmissive object plane modulation, transmissive image plane modulation, reflective object plane modulation, and reflective image plane modulation, according to the type of scene to be detected and the modulation mode.

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

[0045]

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

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

[0048] Compared with the prior art, the present invention has the following beneficial effects: the present invention combines single-pixel detection technology with artificial intelligence technology by designing an optoelectronic hybrid neural network, uses a rotating disk to achieve efficient encoding of object information by the optical end network, and then uses the electrical end network to achieve high-accuracy target recognition from the single-pixel detection signal. Compared with the traditional target recognition technology based on two-dimensional images (pure electronics neural network), the present invention can avoid recording the target object image, reduce the recording of redundant information, improve the efficiency of information acquisition, and can be used for the recognition of high-speed moving objects; compared with the target recognition technology based on all-optical neural network, the present invention can take into account the advantages of optical networks and electrical networks, and can achieve a target recognition system with high accuracy, strong stability and low cost. In addition, due to the advantages of modulation devices and single-pixel detection, the present invention has broad application prospects in target recognition problems in special bands, low light and other scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the optoelectronic hybrid neural network structure provided in an embodiment of the present application.

[0050] Figure 2 It is a schematic diagram of a method for implementing cyclic circular convolution provided in an embodiment of the present application.

[0051] Figure 3 It is a schematic diagram of the method for preparing a rotating disk provided in an embodiment of the present application.

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

[0053] Figure 5 It is a schematic diagram of a target recognition system provided in an embodiment of the present application.

[0054] Figure 6 It is a flow chart of a single-pixel image recognition method based on an optoelectronic hybrid neural network provided in an embodiment of the present application.

[0055] Among them, 1-illumination light source; 2-beam shaping device; 3-rotating disk; 4-microscope objective lens; 5-tube lens; 6-target object; 7-light collecting lens; 8-barrel detector; 9-power supply; 10-motor; 11-data acquisition card; 12-computer. DETAILED DESCRIPTION

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

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

[0058] Step S1: Build a single-pixel optoelectronic hybrid neural network. Figure 1 As shown in the figure, the optical end network of the optoelectronic hybrid neural network realizes the single-pixel detection process of cyclic encoding, and the electrical end realizes the classification of objects from the intensity information of single-pixel detection. Specifically, the input data is The optical end network is R ω in is the circular convolution kernel (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 passed through R ω Single-layer circular convolution processing obtains the light intensity fluctuation I

[0059]

[0060] Here, circular convolution is indirectly implemented through linear convolution. The specific implementation process is as follows Figure 2 As shown, the input image x is first translated and zero-filled to obtain the convolutional layer input Then convolve with ω, we get This is the output of the optical end neural network. In order to facilitate the subsequent use of the rotating disk to implement circular convolution, 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 get the complete encoding mask By ω f From left to right, we can get A coding mask of size M×N. Considering that the coding mask cannot occupy a whole circle, Figure 2 As shown, by duplicating the coding mask and connecting end to end, the space of the rotating disk can be fully utilized to realize cyclic coding.

[0061] For the electrical end network structure, such as Figure 1 As shown, the electrical end neural network takes I as input, and after being processed by the standardization layer, the physical guidance decoding layer, the image recognition network layer, etc., the output is the recognition result.

[0062] The normalization layer processes I as:

[0063]

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

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

[0066]

[0067] in ω i From ωf Intercept from the middle, move one pixel from left to right and take a coding mask of size M×N, and get it every i pixels. A total of A mask.

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

[0069] Step S2: Select a training data set. The data set for target recognition is selected according to the task to be recognized. It can be selected from various open source data sets such as MNIST, Fashion-MNIST, ImageNet, CelebA, etc., or a user-made data set. Any set of data in the data set includes a scene image and the corresponding annotated category.

[0070] Step S3: Use the training data set to train the optoelectronic hybrid neural network. Divide the data set selected in step S2 into a training set, a validation set, and a test set in appropriate proportions. Use the training data set to adjust the optical network parameters and the electrical network parameters by stochastic gradient descent to minimize the error between the output of the optoelectronic hybrid neural network and the label data. The training includes multiple epochs, and the performance of the current model on the validation set is used to evaluate whether the training needs to be terminated to prevent overfitting.

[0071] Step S4: Process the encoding device according to the model parameters obtained by training. Based on the training results obtained in step S3, take out the cyclic encoding mask ω in the optical end network f The rotating disk is placed on a rotating disk. The rotating disk is based on a transparent rigid material and coated in a circular range on the base. The coating structure is a cyclic mask structure obtained through training. f .

[0072] The coating process needs to be based on a rectangular matrix ω f Generate a ring structure 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 re-align ω on the ring f The cyclic mask is arranged through multiple integer cycles to fill the entire ring. The cyclic mask is realized by circular convolution, so that masks of different cycles on the ring can be connected end to end. The motor drives the mask to rotate one mask pixel to achieve one refresh, and one rotation cycle can achieve one record 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, and the formula is:

[0074]

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

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

[0077] Step S5: Build a single pixel detection system based on the encoding device. Figure 4 As shown, the single pixel detection system includes: 1-illumination light source; 2-beam shaping device; 3-rotating disk; 4-microscope objective lens; 5-tube lens; 6-target object; 7-light collecting lens; 8-barrel detector; 9-power supply; 10-motor; 11-data acquisition card; 12-computer and other components. The matching structure includes four forms: transmission type object plane modulation, transmission type image plane modulation, reflection type object plane modulation, and reflection type image plane modulation.

[0078] Step S6: Obtain the encoded object information using the constructed single pixel detection system. Figure 4 At the target object (6) shown in the figure, by turning on the light source and the 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 by the single-pixel detector in turn. The motor records a light intensity value when it rotates through a cyclic mask pixel. According to the set sampling rate, the motor rotates through pixels, that is, a set of light intensity fluctuation signals I for target recognition is obtained. The acquisition and recognition signal 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, the sampling rate is 0.1, and the rate at which the single-pixel detection system collects the light intensity fluctuation signal I can reach 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 collects the light intensity fluctuation signal I can be further increased to 31.4KHz.

[0081] Step S7: Use the 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, it is input into the pre-trained image recognition module at the electrical end, and processed by the standardization layer, the physical guidance decoding layer and the image recognition network layer to obtain the prediction result of the scene category.

[0082] Step S8: Repeat steps S6 to S7 for a new scene to be identified. 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 an image recognition module to obtain a prediction result of the scene category.

[0083] The above explanation of the embodiment of the target recognition method based on the optoelectronic hybrid neural network is also applicable to the target recognition device based on the optoelectronic hybrid neural network and will not be repeated here.

[0084] According to a preferred embodiment of the present invention, a target recognition method and device based on an optoelectronic hybrid neural network is proposed. The optical end uses a single-pixel detection method based on circular convolution, and the electrical end uses image recognition technology to predict the object category from light intensity fluctuations. This method does not require pre-acquisition of images, but only requires acquisition of a small amount of information related to image recognition, which facilitates information acquisition, storage and transmission. The circular convolution at the optical end is achieved by a rotating disk, has an ultra-high refresh rate, and has a wide working band. It can quickly obtain light intensity signals for target recognition and can be used for target recognition under high-speed motion, special bands and low-light conditions.

[0085] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A single-pixel target recognition method based on an optoelectronic hybrid neural network, characterized in that: The following steps are involved: S1.1: Build a single-pixel optoelectronic hybrid neural network with an optical end network and an electrical end network; S1.2: Establish a single-pixel target recognition dataset, which includes object images and categories corresponding to the object images; S1.3: training the single-pixel optoelectronic hybrid neural network using the single-pixel target recognition data set; S1.4: Processing a coding device according to the cyclic mask obtained through training to realize coding of the object image; S1.5: Build a single-pixel detection system based on the encoding device to realize the circular convolution of the optical end network on the object image to obtain the intensity fluctuation signal; S1.6: Obtaining the encoded object information using the single pixel detection system; S1.7: Using the single-pixel optoelectronic hybrid neural network trained in step S3, output the category of the object to be detected according to the object image in the single-pixel detection signal; S1.8: For a new scene to be identified, repeat steps S6 to S7.

2. The single-pixel target recognition method according to claim 1, characterized in that: The single-pixel optoelectronic hybrid neural network is a self-encoding architecture, the optical end network is the encoding layer, the electrical end network is the decoding layer, and the bottleneck layer is the optical end network output, that is, the electrical end network input.

3. The single-pixel target recognition method according to claim 2, characterized in that: The optical end network is a single-pixel detection technology based on a cyclic mask, which takes an object image as input and a light intensity fluctuation signal as output, and is used to implement the design of a coding mask for a single-pixel detection system; the electrical end network predicts the object category from the light intensity fluctuation signal output by the optical end network, and includes a preprocessing layer, a physical guided decoding layer, and an image recognition network layer, and is used to realize target recognition from a single-pixel detection signal.

4. The single-pixel target recognition method according to claim 3, characterized in that: The recurrence mask, i.e., the encoding matrix of the single-pixel detection system, can be adaptively optimized during the training process, and its dimension is determined by the image size and the sampling rate.

5. The single-pixel target recognition method according to claim 3, characterized in that: The preprocessing layer standardizes the single-pixel detection signal so that the subsequent recognition network pays attention to the fluctuation of the single-pixel detection signal, making it suitable for target recognition under different light levels; the physical-guided decoding layer uses the cyclic coding mask in the optical end network to perform intensity differential association operation with the standardized single-pixel detection signal. The operation result can be regarded as the extraction of scene image features. The physics used in the process includes cyclic coding mask (forward physical model) and intensity differential association operation rules (image feature reconstruction method). This process converts the one-dimensional detection signal into a two-dimensional image feature; the image recognition network layer calculates and outputs the object category based on the image features extracted by the physical-guided decoding layer.

6. The single-pixel target recognition method according to claim 1, characterized in that: S1.3: using the single-pixel target recognition data set to train the single-pixel optoelectronic hybrid neural network, specifically: S6.1 divide the single-pixel object recognition dataset into a training set, a validation set, and a test set; each set of data in the dataset includes an image and a category corresponding to the image; S6.2 using the training set, selecting some data pairs in the training set each time, and using the images in the data as inputs of the optoelectronic hybrid neural network, and the corresponding categories as labels corresponding to the inputs; S6.3 The input image is processed by the optoelectronic hybrid neural network to obtain the predicted category, and the predicted category is compared with the true label to obtain the loss function; the gradient of each learnable parameter of the optoelectronic hybrid neural network is calculated according to the back propagation algorithm, and then the parameters are updated using the gradient descent method; binary constraints are introduced into the parameters of the optical end network, and the update of its parameters can only be selected between 0 and 1. S6.4 After each round of training, 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; Test on the test set, and use its performance 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 it does not meet the requirements (the model's recognition accuracy for the test set images is too low), further adjust the network hyperparameters and then train again. The hyperparameters include the given sampling rate, learning rate, and network structure; S6.6 Repeat the above process. When the performance of the trained model on the test set meets the requirements, the current trained network model is regarded as the final training result and the training is completed.

7. The single-pixel target recognition method according to claim 6, characterized in that: During the training process, the learnable parameters of the optical end network are binarized to realize binary cyclic mask encoding; the cyclic mask obtained after the optical end network is trained is realized on the hardware by rotating the disk, and multiple groups of cyclic masks are connected end to end to cover a circle of rings on the disk, and rings with different radii are used to realize the cyclic masks obtained by training under different conditions.

8. The single-pixel target recognition method according to claim 1, characterized in that: The encoding device is a rotating disk, which has a rigid transparent material as a substrate and is coated in a circular range on the substrate. The coating structure is a trained circular mask structure. The rotation of the disk driven by a motor can realize the encoding process of single-pixel detection based on the circular mask.

9. The single-pixel target recognition method according to claim 8, characterized in that: At different radii of the rotating disk, a cyclic mask structure trained under different conditions is coated to obtain multiple concentric rings. By selecting different rings to illuminate the object, the same rotating disk can be used to achieve encoding of the target object under different conditions; The cyclic mask is arranged through multiple integer cycles to fill the entire circle. The cyclic mask is implemented through circular convolution, so that masks of different cycles on the circle can be connected end to end. The motor drives the mask to rotate one mask pixel to achieve one refresh, and one rotation cycle can achieve one record of the signal required for target recognition.

10. The single-pixel target recognition method according to claim 1, characterized in that: The single-pixel detection system includes a power supply, a light source, a data acquisition card, a single-pixel detector, a computer, an encoding device, a motor, a lens, a supporting structure and an aperture. The encoding device is a rotating disk used to realize cyclic encoding of a target object. 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 a computer program; the encoding device is connected to the motor; the lens and the aperture are placed in the optical path to shape the light beam; for the object plane encoding scheme, the light source is modulated by the encoding device to form structured light to illuminate the object; for the image plane encoding scheme, the echo is modulated by the encoding device after the light source illuminates the object; and the supporting structure fixes the components.

11. The single-pixel target recognition method according to claim 7, characterized in that: The single-pixel detection system can image transmissive targets and reflective targets, and the encoding of objects can be on the object plane (structured illumination) or on the image plane (structured detection); the single-pixel detection system obtains target object information by relying on the synchronization of the motor and the single-pixel detector. The position of motor rotation determines the pattern of the current 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 end neural network.

12. The single-pixel target recognition method according to claim 1, characterized in that: The single pixel detection system in step S1.5 comprises: A lighting module, used for lighting the scene to be tested; A control module for controlling the synchronization of the rotary motor and the single pixel detector; A modulation module is connected to the control module and includes a rotating disk and a motor; the rotating disk is based on a transparent rigid material, and a trained cyclic mask is plated on the ring of the base; the motor drives the rotating disk to refresh the coding mask for cyclic coding single pixel detection every time the rotating disk rotates through a mask pixel size; an adjustable adapter is used between the rotating disk and the motor to align the center of the ring with the rotation center of the motor and to make the plane of the disk perpendicular to the motor shaft; The detection module includes a beam collection system, a single-pixel detector, and a data acquisition card, and is used to collect and obtain the modulated light intensity fluctuation signal; The calculation module is connected to the detection module and is used to predict the category of the target scene to be measured from the light intensity fluctuation signal obtained by the detection module.

13. The single-pixel target recognition method according to claim 9, characterized in that: The refresh rate S of the modulation module satisfies the following conditions: Where D is the diameter of the ring mask, T is the motor rotation speed, and P is the pixel size of the mask structure; The target recognition rate of the single-pixel detection system is determined by the refresh rate S and the sampling number The ratio is determined by .

Images