Imaging-free target detection method and system based on single-pixel measurement

Abstract

This invention discloses an imaging-free target detection method and system based on single-pixel measurement, relating to the fields of computational imaging and deep learning technologies. The method includes: training an imaging-free target detection network and a single-pixel modulation module; the single-pixel detection system acquires a one-dimensional measurement array of the current observation area, transmits it to the trained imaging-free target detection network, and acquires the target confidence score and predicted bounding box of the one-dimensional measurement array; based on the target confidence score, it determines whether a target exists in the current observation area; if it exists, the target's location is identified based on the predicted bounding box; if it does not exist, the current observation area is changed, and a new target detection is performed. This invention also discloses an imaging-free target detection system using the above method. Through the above technical solution, this invention establishes an end-to-end mapping from single-pixel measurement data to target detection results without performing two-dimensional image reconstruction, shortening the processing link.

Description

Technical Field

[0001] This invention belongs to the field of computational imaging and deep learning technology, specifically relating to an imaging-free target detection method and system based on single-pixel measurement. Background Technology

[0002] Traditional photoelectric detection imaging methods primarily rely on array detectors such as CCD and CMOS to acquire images of the target scene, followed by image processing or deep learning algorithms to complete the target detection task. While this technical approach is mature and intuitive, its limitations are becoming increasingly apparent as detection tasks expand to longer distances, larger areas, and weaker signals. First, in civilian remote sensing scenarios such as maritime target detection and large-area agricultural remote sensing, traditional detectors typically require large-scale array scanning of vast areas. The massive amount of data puts significant pressure on front-end acquisition, data transmission, and back-end processing. Area array detectors are expensive or even impossible to manufacture, limiting the applicable scenarios for the imaging system. Second, in target detection tasks, the information of interest often only occupies a small portion of the scene, and the acquisition, transmission, and processing of a large number of background pixels results in significant resource waste. Finally, the process from image acquisition to target recognition requires a complete image reconstruction and subsequent processing chain, introducing additional computational delays that are difficult to meet the real-time requirements of rapid detection scenarios.

[0003] Single-pixel imaging offers a new technological path to overcome the limitations of traditional photoelectric detection techniques. It involves encoding and modulating the target light field using a spatial light modulator, followed by collecting the total light intensity response using a single-point detector with no spatial resolution, thus encoding the spatial information of the target scene into a one-dimensional measurement sequence. The single-pixel measurement mechanism can acquire compressed measurement information of the target scene under low sampling rate conditions, thereby reducing the amount of data collected and alleviating the pressure on data storage, transmission, and computation. However, most existing single-pixel imaging methods follow an imaging-then-sensing approach: first, compressed measurement data is acquired through a single-pixel measurement mechanism; then, the target image is reconstructed using methods such as correlation imaging, compressed sensing, or deep learning; finally, the reconstructed image is input into a detection algorithm to complete the sensing task. This process increases the processing link length and computational overhead, reducing the efficiency of information utilization between measurement data and detection results.

[0004] Furthermore, in existing imaging-free sensing methods, the design of single-pixel modulation matrices mostly adopts fixed modulation matrices (such as random speckle or Hadamard matrices) or pursues optimization at the mathematical level, without fully considering the compatibility with the non-negative modulation characteristics of spatial light modulators. This makes it difficult to directly load the trained modulation matrix into hardware implementation, which restricts its engineering deployment on resource-constrained platforms in the field of civilian remote sensing.

[0005] In summary, given the significant shortcomings of existing technologies in handling link length, data redundancy, and computational overhead, there is an urgent need to propose an end-to-end target detection method that can simultaneously determine the existence of a target and predict its spatial location directly from single-pixel low-dimensional measurement data without performing two-dimensional image reconstruction. Summary of the Invention

[0006] To address the shortcomings of existing technologies, such as single-pixel imaging methods relying on image reconstruction and imaging-free sensing methods lacking spatial localization capabilities, this invention proposes an imaging-free target detection method and system based on single-pixel measurement. This method directly establishes an end-to-end mapping from single-pixel measurement data to target detection results without performing two-dimensional image reconstruction. A multi-layer perceptron (MLP) feature extraction module extracts abstract features related to target detection layer by layer. A dual-branch detection module simultaneously outputs target confidence and predicted bounding boxes. This effectively shortens the processing chain from data acquisition to detection result output, reducing data redundancy and computational overhead.

[0007] On one hand, the present invention provides an imaging-free target detection method based on single-pixel measurement, comprising the following steps:

[0008] S1: Construct a sample set to train an imaging-free target detection network and a single-pixel modulation module;

[0009] S2: The single-pixel detection system acquires a one-dimensional measurement array of the current observation area;

[0010] S3: Transmit the one-dimensional measurement array to the trained imaging-free target detection network to obtain the target confidence and predicted bounding box of the one-dimensional measurement array;

[0011] S4: The detection output module determines whether there is a target in the current observation area based on the target confidence level. If there is a target, the target's location is marked according to the predicted bounding box. If there is no target, the current observation area is changed, and the process proceeds to step S2.

[0012] The imaging-free target detection network includes a multilayer perceptron feature extraction module and a dual-branch detection module. Step S3 specifically includes:

[0013] S31: The multilayer perceptron feature extraction module performs nonlinear feature mapping on the one-dimensional measurement array to obtain the target feature vector;

[0014] S32: The dual-branch detection module extracts features from the target feature vector and outputs the target confidence score and the predicted bounding box.

[0015] On the other hand, the present invention provides an imaging-free target detection system based on single-pixel measurement, which uses the above-mentioned method to detect targets in a scene. The system includes:

[0016] A single-pixel detection system is used to acquire a one-dimensional measurement array of the current observation area;

[0017] An imaging-free target detection network includes a multilayer perceptron feature extraction module and a dual-branch detection module. The multilayer perceptron feature extraction module is used to perform nonlinear feature mapping on the one-dimensional measurement array to obtain the target feature vector. The dual-branch detection module is used to simultaneously calculate the target confidence and predict the bounding box based on the target feature vector.

[0018] The detection output module is used to identify the target in the current observation area based on the target confidence level and the predicted bounding box.

[0019] The present invention achieves the following beneficial effects through the above technical solution:

[0020] (1) This invention compresses high-dimensional training images into a one-dimensional measurement array using a single-pixel modulation module, and directly outputs the target confidence and predicted bounding box through a dual-branch detection module. No two-dimensional image reconstruction is required in the entire processing chain, reducing redundant computation and intermediate data storage, and shortening the processing chain from data acquisition to detection result output. This makes it suitable for fast detection scenarios with high real-time requirements. Through the collaborative design of the confidence prediction branch and the bounding box regression branch in the dual-branch detection module, direct regression from the one-dimensional measurement array to the target spatial location is achieved, as well as spatial localization of the target detection task in the field of imaging-free perception.

[0021] (2) This invention employs a two-stage training strategy to train the imaging-free target detection network: the first stage jointly optimizes the modulation matrix and the back-end detection network, enabling the front-end measurement process to adaptively retain effective information related to target existence judgment and spatial location prediction based on the loss feedback of the detection task, rather than focusing on image reconstruction fidelity as in traditional methods; the second stage freezes the modulation matrix and fine-tunes the back-end network, allowing the model to further adapt to the target detection task under a fixed measurement mode. Experimental results show that this training strategy effectively improves the model's detection precision, detection recall, and localization success rate, verifying the effectiveness of trainable modulation matrices in imaging-free target detection tasks.

[0022] (3) By simulating the non-negative transmittance or reflectance physical characteristics of the spatial light modulator, the modulation method of the modulation matrix is ​​matched with the non-negative physical modulation characteristics of the spatial light modulator, ensuring that the modulation matrix obtained by training can be directly loaded into the spatial light modulator hardware implementation. This effectively overcomes the problem of training and deployment disconnect that is common in existing imaging-free sensing methods, and provides a feasible path for engineering deployment on resource-constrained platforms in the field of civilian remote sensing.

[0023] (4) For hardware scenarios that only support binary modulation, this invention further provides an optional implementation of the binarized modulation matrix, which achieves end-to-end training through a sign function and a forward pass-through estimator. Compared to the sensitivity of image reconstruction tasks to quantization errors, target detection tasks have a higher tolerance for binarized quantization. Even with a binarized modulation matrix, it can still maintain target presence judgment and coarse localization capabilities similar to continuous value modulation, thus providing flexible choices for deployment platforms with different hardware capabilities. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating the steps of the imaging-free target detection method based on single-pixel measurement according to an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram of the structure of a single-pixel modulation module according to an embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of the network module structure during the training phase of an embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram of the structure of the MLP feature extraction module according to an embodiment of the present invention;

[0029] Figure 5 This is a schematic diagram of the dual-branch detection module according to an embodiment of the present invention;

[0030] Figure 6 This is a comparison chart of experimental results for different modulation matrices in the experiments of this invention;

[0031] Figure 7 This is a graph showing the performance changes of the model under different sampling rates in the experiments of this invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Traditional imaging-then-sensing techniques require acquiring a complete image using an array detector before employing image processing or deep learning algorithms to perform target detection. This approach has limitations in handling link length, data redundancy, and computational overhead. The core technical problem this invention aims to solve is: how to simultaneously determine the existence and spatial location of a target directly from single-pixel low-dimensional measurement data without performing two-dimensional image reconstruction. Solving this problem requires answering two key questions: first, why can one-dimensional measurement data contain information about the target's existence and spatial location? Second, what network structure can effectively extract this information from one-dimensional measurement data that has lost its spatial structure? The following detailed embodiments illustrate the technical concept of this invention.

[0034] Example 1

[0035] In one embodiment, the present invention proposes an imaging-free target detection method based on single-pixel measurement. This method directly establishes an end-to-end mapping from single-pixel measurement data to target detection results without performing two-dimensional image reconstruction, avoiding computational redundancy and link delays caused by image reconstruction in traditional processes.

[0036] Please see the appendix Figure 1 , Figure 1 This is a flowchart illustrating the steps of the imaging-free target detection method based on single-pixel measurement according to an embodiment of the present invention. As shown in the figure, the method includes:

[0037] S1: Construct a sample set to train an imaging-free target detection network and a single-pixel modulation module;

[0038] S2: The single-pixel detection system acquires a one-dimensional measurement array of the current observation area;

[0039] S3: Transmit the one-dimensional measurement array to the trained imaging-free target detection network to obtain the target confidence and predicted bounding box of the one-dimensional measurement array;

[0040] S4: The detection output module determines whether the target exists in the current observation area based on the target confidence level. If it exists, the target's location is marked according to the predicted bounding box. If it does not exist, the current observation area is changed, and the process proceeds to step S2.

[0041] The specific implementation methods for each step are described in detail below.

[0042] I. The specific steps of S1 include:

[0043] S11: Construct a sample set, which includes target samples and non-target samples; the sample structure of the sample set includes training images, confidence labels, and bounding box labels;

[0044] A target sample is an image sample constructed from the original image that contains one and only one target. The confidence label for a target sample is set to 1, and it also includes a bounding box label. The specific construction method for target samples is as follows:

[0045] A target is selected from the original image as the anchor point. The cropping position is randomly determined, generating a cropping region (e.g., a 256×256 pixel region) that completely contains the anchor point. This cropping region does not exceed the boundary coordinates of the original image, and there is no overlap between the cropping region and the bounding box labels of any non-anchor point targets in the original image, ensuring that the cropped region contains only one complete target. The cropping region is then scaled to a preset pixel size (e.g., 64×64 pixels). For example, scaling from 256×256 pixels to 64×64 pixels corresponds to a scaling ratio of [missing value]. The bounding box labels of the anchor point targets are then synchronously transformed using the same scaling ratio. The coordinate transformation method is as follows:

[0046] ; ;

[0047] ; ;

[0048] in, Here are the coordinates of the top-left corner of the cropped area in the original image's coordinate system. Let be the bounding box coordinates of the anchor point target in the original image coordinate system, where This represents the coordinates of the top-left corner of the bounding box. Indicates the coordinates of the bottom right corner of the bounding box; Let be the coordinates of the transformed bounding box, where This represents the coordinates of the top-left corner of the transformed bounding box. This represents the coordinates of the bottom right corner of the transformed bounding box.

[0049] Target-free samples are image samples constructed from the original image that do not contain any targets. The confidence label of a target-free sample is set to 0, and its bounding box label is set to invalid. The specific construction method for target-free samples is as follows: a cropped region (e.g., a 64×64 pixel region) is randomly selected from the original image. If the cropped region does not overlap with any bounding box labels in the original image, it is saved as a target-free sample.

[0050] Through the above construction method, the multi-target samples that may exist in the original dataset are transformed into single-target samples that are more suitable for training the network of this invention. Samples with targets are positive samples, used to train the model to learn the measurement features and location distribution under the condition that the target is present; samples without targets are negative samples, used to train the model to learn the measurement features under the condition that there is no target background. The combined effect of positive and negative samples enables the model to have both target detection capability and background suppression capability, thereby reducing the false alarm rate.

[0051] S12: The training image is converted into the one-dimensional measurement array by the single-pixel modulation module;

[0052] The single-pixel modulation module of this invention is used to simulate the spatial coding and compressed sampling process of a single-pixel detection system during the training phase, with the aim of generating a one-dimensional measurement array corresponding to the sample set.

[0053] Please see the appendix Figure 2 , Figure 2 This is a schematic diagram of the single-pixel modulation module according to an embodiment of the present invention. As shown in the figure, in this embodiment, the single-pixel modulation module includes an image conversion module and a single-pixel modulation matrix. The steps of using the single-pixel modulation module to convert the training image specifically include: the image conversion module unfolds the training image, converts it into one-dimensional data in equal quantities, and performs pixel normalization processing; the single-pixel modulation matrix compresses and samples the normalized one-dimensional data to obtain the original one-dimensional measurement array.

[0054] For example, first, the size is Image Flattened into a one-dimensional image array in pixel order The length of the flattened vector is and the Normalizing to a range of 0 to 1 yields a normalized one-dimensional image array:

[0055] ;

[0056] in, Represents a flattened one-dimensional image array The first in pixel value, Represents the normalized i-th Each pixel value; This represents the maximum grayscale value determined by the pixel bit depth of the trained image. The core of this module is a trained single-pixel modulation matrix. Single-pixel modulation matrix Each row corresponds to a spatial light modulation pattern, and each element represents the contribution weight of that spatial location to the detector response.

[0057] For example, for an 8-bit grayscale image, the corresponding After pixel normalization, the following is obtained Then, a single-pixel modulation matrix with dimensions of 1024×4096 is used. (Sampling rate selection) Compressed measurements are performed, outputting a 1024-dimensional measurement vector. This refers to the original one-dimensional measurement array, and the entire processing does not involve any two-dimensional image reconstruction operations.

[0058] In an optional embodiment, to improve the stability of the numerical distribution of the measurement vector (the original one-dimensional measurement array), a batch normalization layer and a ReLU activation function are added to the single-pixel modulation module to perform batch normalization and nonlinear activation on the original one-dimensional measurement array, and output the one-dimensional measurement array.

[0059] The batch normalization layer standardizes the measurement vector (the original one-dimensional measurement array) for each batch of samples:

[0060] ;

[0061] in, and These represent the current batch number. The mean and variance of each measurement value To prevent division by zero, a ReLU activation function is used after batch normalization to introduce nonlinearity, enabling the network to extract nonlinear features.

[0062] The image of the scene to be detected is a two-dimensional light intensity distribution. To obtain spatial information of the scene to be detected, a series of modulation patterns with spatial resolution need to be generated. Each modulation pattern is projected onto the target or placed in front of the detector. After the light field is spatially modulated, it is reflected by the target and collected by a single-pixel detector to obtain a series of light intensity measurements. Vectorization operators are then used... Two-dimensional image Flattened into a one-dimensional image array in pixel order , The length of the one-dimensional image array after the image is flattened. Arrange the modulation pattern sequence into a single-pixel modulation matrix. The entire measurement process can then be represented as a linear system:

[0063] ;

[0064] in, For a one-dimensional measurement array, This represents the number of measurements taken per pixel, i.e., the length of the one-dimensional measurement array. Each measurement value... This can be represented as a weighted sum of the image and a certain modulation pattern:

[0065] ;

[0066] in, Indicates the first In the modulation pattern, the first Weights at each pixel position ; Indicates the first Each pixel value. This formula indicates that the actual single-pixel detector during detection does not directly acquire the spatial distribution of each pixel in the image, but rather obtains the overall response results of the entire image under different modulation modes. The spatial location information of the target is implicitly encoded in the combination relationship between different measurements. This is the physical basis for the present invention to directly extract the target's existence information and spatial location information from a one-dimensional measurement array, thereby bypassing the traditional image reconstruction steps.

[0067] After single-pixel modulation, the one-dimensional measurement array no longer possesses the clear local spatial structure of a two-dimensional image. The concept of "adjacent pixels" no longer exists in the measurements, and all measurements are globally coupled. Convolutional neural networks, which rely on local receptive fields, struggle to perform well with this type of data. In contrast, the fully connected structure of an MLP can model the global correlation between any two measurement dimensions without bias. Therefore, this invention chooses an MLP as the main structure of the feature extraction network.

[0068] The trainable characteristics of the single-pixel modulation matrix: Unlike the fixed modulation matrix used in traditional single-pixel imaging, such as random speckle matrix, Hadamard matrix or Fourier matrix, the modulation matrix A of this invention is updated by backpropagation based on the loss of the detection task during training, so that the front-end measurement process can adaptively prioritize the retention of effective information related to the target detection task, rather than using image reconstruction fidelity as the optimization target as in traditional methods.

[0069] The degree of compression in measurement is controlled by the sampling rate, which is defined as follows: The number of rows in the single-pixel modulation matrix and the length of the one-dimensional measurement array are both equal to the number of single-pixel measurements. .

[0070] The sampling rate is set between 12.5% ​​and 50% to ensure that the ability to determine the existence of the target and locate the target is maintained under undersampling conditions, while achieving a balance between the amount of measurement information and the data processing cost. Preferably, the sampling rate is set to... This approach achieves an optimal balance between measurement information, model generalization ability, and data processing cost. Subsequent experimental data will further demonstrate that target detection performance (detection precision, detection recall, and localization success rate) exhibits a non-monotonic trend. In fact, performance decreases rather than increases after the sampling rate reaches a certain level, indicating the counterintuitive rule that "excessively high sampling rates lead to overfitting."

[0071] In one exemplary embodiment, when , hour, .

[0072] S13: Input the one-dimensional measurement array into the imaging-free target detection network to obtain the target confidence and the predicted bounding box;

[0073] Please see the appendix Figure 3 , Figure 3 This is a schematic diagram of the network module structure during the training phase of an embodiment of the present invention. As shown in the diagram, the one-dimensional measurement array output by the single-pixel modulation module is input to the imaging-free target detection network. The imaging-free target detection network includes a multilayer perceptron feature extraction module and a detection output module. The multilayer perceptron feature extraction module is used to perform nonlinear feature mapping on the one-dimensional measurement array to obtain the target feature vector, corresponding to the attached... Figure 1 Step S31. The dual-branch detection module is used to extract features from the target feature vector and output the target confidence score and the predicted bounding box, corresponding to the attached... Figure 1 Step S32 in the process.

[0074] (1) Multilayer perceptron feature extraction module

[0075] The multilayer perceptron is a typical feedforward neural network model, and its basic working principle is as follows:

[0076] Let the input of the g-th layer be the output a(g-1) of the previous layer, the weight matrix be W(g), and the bias vector be b(g). Then, this layer first performs a linear transformation:

[0077] z(g) = W(g)a(g-1) + b(g);

[0078] Where z(g) represents the linear output of the g-th layer.

[0079] The linear output is then passed through the activation function f(·) to obtain the final output of this layer:

[0080] a(g) = fz(g);

[0081] If the network contains a total of L layers, then the entire multilayer perceptron can be represented as a composition of multiple activation functions:

[0082] y=Fx;θ=f(L)(W(L)f(L-1)...f(1)W(1)x+b(1)...+b(L));

[0083] Where x represents the input data, y represents the network's predicted output, and θ represents all trainable parameters in the network, including the weight matrices and bias vectors of each layer. The activation function introduces nonlinear modeling capability to the neural network. f(1)(·)=max(0,·) represents the ReLU function of the g-th layer, which is adopted in this invention because of its simple form, high computational efficiency, and ability to alleviate the gradient vanishing problem to some extent.

[0084] Please see the appendix Figure 4 , Figure 4 This is a schematic diagram of the MLP feature extraction module according to an embodiment of the present invention. As shown in the figure, based on the network structure principle of the MLP described above, the multilayer perceptron feature extraction module employs several fully connected layers connected end-to-end, with no convolution, pooling, or attention mechanisms between the fully connected modules. By setting the dimension of the output vector of each fully connected module to decrease progressively, the one-dimensional measurement array is compressed layer by layer; in each fully connected module, the following steps are performed:

[0085] The fully connected layer performs an affine mapping on the input vector to extract a low-dimensional vector.

[0086] The activation layer performs a non-linear transformation on the low-dimensional vector to extract feature vectors; the activation layer is activated using the ReLU function.

[0087] During training, the Dropout layer randomly sets some elements of the feature vector to zero, outputting this as the output vector to the next fully connected module. This random deactivation mechanism prevents the network from overly relying on certain fixed features, thereby improving its generalization ability.

[0088] During testing, the Dropout layer directly uses the feature vector as the output vector and outputs it to the next fully connected module.

[0089] The dimension of the input vector of each of the fully connected modules is greater than the dimension of the output vector;

[0090] In this embodiment of the invention, the number of fully connected modules is set to 4, and the random dropout rate of the Dropout layer is set to 0.5. In one or other embodiments of the invention, the number of fully connected layers and the specific dimensions of each layer can be adjusted according to actual task requirements.

[0091] (2) Detection output module

[0092] The detection output module of this embodiment includes a confidence prediction branch and a bounding box regression branch, for a total of two branches. The two branches share the input target feature vector, and independently extract the target confidence and predict the bounding box from the target feature vector.

[0093] The technical significance of the dual-branch design is that the confidence prediction branch is essentially a binary classifier used to determine whether a target exists in the current sample; the bounding box regression branch is a continuous coordinate regressor used to further infer the target's location range in the original image space after confirming its existence. The two branches share the feature vector output by the MLP module but undertake different sub-tasks, forming a collaborative output relationship between "target presence / absence judgment" and "target location prediction." This invention, through the collaborative design of the confidence prediction branch and the bounding box regression branch, achieves direct regression from a one-dimensional measurement array to the target's spatial location in the field of image-free perception, overcoming the technical limitation of existing image-free perception methods that can only output category labels and cannot perform spatial localization.

[0094] Preferably, the calculation process for the target confidence level is as follows:

[0095] ;

[0096] The calculation process for extracting the predicted bounding box is as follows:

[0097] ;

[0098] in, The target confidence level, Let H be the predicted bounding box, Sigmoid be the activation function, h be the target feature vector, and W be the target bounding box. conf b is the weight matrix for the confidence prediction branch. conf W is the bias vector for the confidence prediction branch. reg Let b be the weight matrix of the bounding box regression branch. reg This is the bias vector for the bounding box regression branch;

[0099] The target confidence level is set as a one-dimensional vector, and the predicted bounding box is set as a four-dimensional vector.

[0100] In particular, in order to enable the predicted bounding box and target confidence values ​​output by the dual-branch detection module to be adjusted independently, the present invention performs independent dimensionality adjustment processing on the target feature vector in the confidence prediction branch and the bounding box regression branch, respectively.

[0101] Please see the appendix Figure 5 , Figure 5 This is a schematic diagram of the dual-branch detection module according to an embodiment of the present invention. As shown in the figure, in this embodiment, the two branches of the dual-branch detection module share the 64-dimensional shared feature vector output by the MLP feature extraction module. The confidence prediction branch consists of a fully connected layer FC1 with an output dimension of 128, a fully connected layer FC2 with an output dimension of 64, and a linear layer with an output dimension of 1, outputting the target confidence score. The bounding box regression branch consists of a fully connected layer FC1 with a 128-dimensional output, a fully connected layer FC2 with a 64-dimensional output, and a linear layer with a 4-dimensional output, which outputs the predicted bounding boxes. Both branches' output layers use the Sigmoid activation function, constraining the output values ​​between 0 and 1, making the predicted bounding boxes normalized coordinates relative to the training image size. Here, FC stands for Fully Connected Layer, and Linear represents a linear output layer.

[0102] S14: Calculate the loss function based on the target confidence and the predicted bounding box. If the loss function meets the preset conditions, end the training. If the loss function does not meet the preset conditions, jointly adjust the parameters of the single-pixel modulation module and the imaging-free target detection network, and return to S12.

[0103] S15: Obtain the trained single-pixel modulation module and the trained imaging-free target detection network.

[0104] As can be seen, steps S14 and S15 mainly train the imaging-free target detection network and the single-pixel modulation module through a loss function. The calculation steps for the loss function include:

[0105] (1) Calculate the confidence loss We employ binary cross-entropy loss:

[0106] ;

[0107] in, The target confidence level, The confidence level label is...

[0108] (2) Calculate the bounding box regression loss :

[0109] ;

[0110] in, The intersection-union ratio (IoU) of the predicted bounding box and the bounding box label is given. The square of the Euclidean distance between the predicted bounding box and the center point of the bounding box label; The square of the diagonal length of the minimum bounding rectangle of the predicted bounding box and the bounding box label; This is a consistent aspect ratio item; These are the weighting coefficients.

[0111] (3) Calculate the loss of the target sample. :

[0112] ;

[0113] in, The coordinate loss weights are used.

[0114] (4) Calculate the loss of the targetless sample. :

[0115] ;

[0116] in, The weights are for loss without a target.

[0117] (5) Calculate the loss function L for training in the same batch:

[0118] ;

[0119] in, Indicates batch size; Indicates the first sample in the sample set The total loss for the nth sample; when the nth sample When each sample is the target sample. When the first When the sample is the target sample, .

[0120] Compared to traditional L1 or L2 coordinate losses, CIoU loss not only considers the numerical difference between the predicted and ground truth bounding box coordinates, but also comprehensively considers intersection-over-union ratio, center distance, and aspect ratio consistency, making it more suitable for bounding box regression optimization in object detection. (Here, bounding box labels...) The supervisory signal is the predicted bounding box, which is the model output and is a 4-dimensional coordinate vector. After introducing CIoU loss, the network will tend to optimize the predicted box that is closer to the center point and has a more reasonable aspect ratio, which is beneficial to improving the quality of bounding box localization.

[0121] In one exemplary embodiment, setting It is 10.0.

[0122] The total loss for samples without a target is calculated by multiplying the confidence loss by the weight of the loss without a target. This is the weight for the targetless loss, used to balance the impact of positive and negative samples on the training process; in this embodiment, it is set to 0.8. Since targetless samples do not have bounding boxes, the bounding box regression loss is not included in the total loss calculation for targetless samples.

[0123] The joint adjustment process in step S14 includes the parameter adjustment and optimization process of the target detection network and the single pixel module.

[0124] Preferably, the embodiments of the present invention employ a two-stage training strategy:

[0125] The first stage is the joint training stage. All trainable parameters in the imaging-free target detection network participate in joint optimization, enabling the single-pixel modulation matrix to be adaptively updated according to the detection loss. The trainable parameters include at least the single-pixel modulation matrix of the single-pixel modulation module, the weights and biases of each layer of the MLP feature extraction module, and the weights and biases of the dual-branch detection module.

[0126] Specifically, the single-pixel modulation matrix is ​​updated via gradient backpropagation based on the weighted sum of the confidence loss and the bounding box regression loss, enabling it to adaptively evolve from an initial random modulation mode to an optimized measurement mode for the target detection task. The role of the joint training phase is that the single-pixel modulation matrix is ​​no longer designed to preserve image reconstruction information, but rather to serve the target detection task, thereby learning a measurement representation more suitable for the current target detection task. Through joint training, the modulation module can learn the most suitable measurement representation for the current target detection task, improving the efficiency of subsequent networks in utilizing target features.

[0127] The second stage is the modulation freeze training stage. The single-pixel modulation matrix of the single-pixel modulation module is frozen, and only the multilayer perceptron feature extraction module and the dual-branch detection module are trained to adapt them to a fixed measurement pattern. Specifically, the single-pixel modulation matrix of the single-pixel modulation module and the statistical parameters of the batch normalization layer are frozen, and only the weights and biases of the fully connected layers of the multilayer perceptron feature extraction module and the dual-branch detection module are trained.

[0128] The purpose of setting up the second stage is twofold: First, it can improve the adaptability of the back-end detection network to fixed measurement vectors. Under the condition that the modulation module is frozen, the back-end network needs to further mine the target information in the measurement data without changing the measurement method. Second, it is closer to the deployment method of actual single-pixel detection systems. In real hardware systems, once the modulation pattern is determined, it is usually not advisable to change it frequently. Continuing to train the back-end network after freezing the single-pixel modulation matrix can simulate the optimization process of the detection network under a fixed modulation pattern, which has engineering value.

[0129] In one embodiment, the first stage is trained using a first learning rate, and the second stage is trained using a second learning rate, wherein the second learning rate is less than the first learning rate.

[0130] In one optional embodiment, the first learning rate is set to 0.001, and the number of training rounds in the first stage is set to 150. A larger initial learning rate helps the model quickly search for optimal parameter regions in the early stages of training, enabling the modulation module and the detection network to form an effective cooperation as soon as possible.

[0131] In one optional embodiment, the second learning rate is set to 0.0001, and the number of training epochs in the second stage is set to 100. Compared to the first stage, the second stage uses a smaller learning rate, the purpose of which is to make fine adjustments based on the existing model and avoid large parameter updates from destroying the effective feature representations already formed in the first stage.

[0132] In one embodiment, the modulation mode of the single-pixel modulation matrix simulates the non-negative transmittance or reflectance physical characteristics of a spatial light modulator, so that the modulation mode of the single-pixel modulation matrix during the single-pixel modulation process matches the physical modulation characteristics of the spatial light modulator.

[0133] Specifically, all elements of the single-pixel modulation matrix are non-negative; the initial values ​​of the elements of the single-pixel modulation matrix are set to a uniform distribution between 0 and 1. Spatial light modulators (such as digital micromirror devices) only support non-negative modulation (0 / 1 or 0~255 grayscale). The non-negativity constraint ensures that the single-pixel modulation matrix is ​​compatible with physical hardware from the initial design stage, ensuring that "mathematical optimality" and "hardware feasibility" are unified, and avoiding the problem that the trained single-pixel modulation matrix cannot be deployed on the spatial light modulator.

[0134] In an alternative embodiment, to accommodate some spatial light modulators that only support binary modulation (e.g., digital micromirror devices), a binarization constraint may be applied to the single-pixel modulation matrix.

[0135] The single-pixel modulation matrix is ​​a binary matrix. In actual single-pixel imaging systems, spatial light modulators mainly achieve binary modulation of the light field by controlling the flip state of micromirrors. The transmittance or reflectance of a single modulation unit usually only has two states: "on" and "off", with corresponding modulation weights of "1" and "0" respectively, which cannot directly achieve continuous grayscale modulation.

[0136] To simulate this physical constraint and improve the physical compatibility between the single-pixel modulation matrix and the actual spatial light modulator, the single-pixel modulation matrix needs to be binarized during the training process of the imaging-free target detection network. Specifically, this includes:

[0137] The trainable parameters corresponding to the binary matrix are stored in the form of a real-valued matrix.

[0138] In forward propagation, the real-valued matrix is ​​projected into the binary matrix using a sign function. The sign function maps real-valued trainable parameters greater than zero in the real-valued matrix to 1, and real-valued trainable parameters less than or equal to zero to 0, thereby simulating the binary physical modulation characteristics (non-negative transmittance or reflectance) of a spatial light modulator. This process can be represented as:

[0139] ;

[0140] in, It is a symbolic function.

[0141] Thus, the single-pixel modulation matrix actually involved in the compression measurement is a binary matrix. This matrix simulates the binary physical modulation characteristics of a spatial light modulator.

[0142] In backpropagation, since the gradient of the sign function is almost everywhere zero, the gradient cannot be backpropagated to the real-valued trainable parameters through the binarization operation. To solve this problem, a pass-through estimator is used to directly pass the gradient of the loss function with respect to the binarized output to the corresponding real-valued trainable parameters, i.e.:

[0143] ;

[0144] Binarization is used in the forward propagation. In computation, the derivative of the binarization operation is ignored during backpropagation, and the gradient is directly passed to the real-valued trainable parameters. This supports end-to-end optimization; the loss function is a normalized result of the sum of the total loss of all target samples and the total loss of non-target samples in a batch, divided by the total number of samples in that batch.

[0145] Through the above binarization design, the single-pixel modulation matrix can directly correspond to a set of binary modulation patterns that can be loaded into the spatial light modulator after training. This effectively avoids information loss when converting continuous weights to binary patterns and improves the transfer capability from simulation training to physical deployment.

[0146] The binarization scheme serves two purposes: firstly, the sign function in forward propagation makes the single-pixel modulation matrix feasible in real-world physical systems; secondly, the pass-through estimator in backpropagation ensures that the single-pixel modulation matrix can still be optimized and updated using gradient descent, allowing the measurement process to adaptively adjust according to the detection task. For target detection tasks, although the binarized single-pixel modulation matrix loses some grayscale detail information, the determination of the target's existence and approximate location prediction mainly rely on the relative response differences between measured values ​​rather than absolute grayscale values. Therefore, in detection tasks, the binarized single-pixel modulation matrix can still maintain a detection performance level similar to continuous value schemes. This high tolerance for binarization quantization is a key characteristic that distinguishes it from image reconstruction tasks.

[0147] II. Specific steps of S2

[0148] After training the neural network model, the imaging-free target detection network can be used directly for target detection. Before that, step S2 is required: using a single-pixel detection system to obtain a one-dimensional measurement array of the current observation area.

[0149] The specific steps include:

[0150] (1) Deploy a single-pixel detection system, including a near-infrared light source, a spatial light modulator, a single-pixel detector, and computing equipment. The system is installed on a lighthouse or a tower along the coast near the port to periodically monitor passing ships in the waterway.

[0151] (2) When the system is working, the spatial light modulator sequentially loads the modulation patterns trained by this invention (for example, 1024 modulation patterns are set, corresponding to a sampling rate of 25%). Each modulation pattern spatially encodes the observation area, and the single-pixel detector synchronously collects the corresponding total light intensity response, thereby obtaining a one-dimensional measurement array. This measurement array is transmitted to the edge computing device for real-time processing via a wireless network.

[0152] III. Specific Steps of S3

[0153] S3: Transmit the one-dimensional measurement array to the trained imaging-free target detection network to obtain the target confidence and predicted bounding box of the one-dimensional measurement array; specifically including: S31: The multilayer perceptron feature extraction module performs nonlinear feature mapping on the one-dimensional measurement array to obtain the target feature vector; S32: The dual-branch detection module extracts features from the target feature vector and simultaneously outputs the target confidence and the predicted bounding box.

[0154] In a preferred embodiment, the plurality of fully connected modules includes a first fully connected module, a second fully connected module, a third fully connected module, and a fourth fully connected module. The dimension of the one-dimensional measurement array is set to 1024; the dimension of the output vector of the first fully connected module is set to 512; the dimension of the output vector of the second fully connected module is set to 256; the dimension of the output vector of the third fully connected module is set to 128; and the dimension of the output vector of the fourth fully connected module is set to 64. Thus, the multilayer perceptron feature extraction module performs layer-by-layer nonlinear transformations on the one-dimensional measurement array to extract feature vectors.

[0155] Specifically, when the sampling rate η=25% and N=4096, the input one-dimensional measurement array has a dimension M=1024. The first layer maps it to 512 dimensions, reducing the feature dimension while retaining the main information. The subsequent three layers progressively compress it to 256, 128, and 64 dimensions, allowing the network to extract more abstract target-related features layer by layer. For example... Figure 4 As shown, the one-dimensional measurement array y is transformed stepwise through four fully connected layers (FC1, FC2, FC3, FC4), with the output feature dimensions of each fully connected layer being 512, 256, 128, and 64 dimensions, respectively. Each fully connected layer is followed by a ReLU activation function and a Dropout layer with a dropout probability of 0.5. There are no convolutional operations, pooling operations, or attention mechanisms between the fully connected layers. The final output is a 64-dimensional shared feature vector h, which serves as the input to the subsequent dual-branch detection module. Here, FC represents a fully connected layer, ReLU represents a rectified linear unit activation function, and Dropout represents a regularization layer with a dropout probability of 0.5.

[0156] In the MLP feature extraction module, higher-dimensional layers are biased towards preserving measurement details, while lower-dimensional layers are biased towards forming task-related representations. The final output feature vector will serve as the shared input to the dual-branch detection module.

[0157] The dual-branch detection module is used to simultaneously output target confidence and predicted bounding boxes. The dual-branch detection module includes a confidence prediction branch and a bounding box regression branch. The two branches share the feature vector output by the MLP feature extraction module, but have independent network parameters and perform different prediction tasks respectively.

[0158] In one embodiment of the present invention, the confidence prediction branch consists of a linear layer and a sigmoid function:

[0159] ;

[0160] in, This represents the feature vector output by the MLP feature extraction module. The target confidence level for prediction; It is the weight matrix of the confidence prediction branch. That is the corresponding bias vector. When The closer it is to 1, the higher the probability that the network believes the target exists in the sample; when... The closer it is to 0, the higher the probability that the network believes the target does not exist in the sample.

[0161] The bounding box regression branch also consists of a linear layer and a sigmoid function:

[0162] ;

[0163] in, The predicted bounding box is the prediction vector that describes the predicted bounding box. For example, when using a horizontal bounding box, the 4-dimensional prediction vector corresponds to the coordinates of the top left corner and the bottom right corner of the predicted bounding box, respectively. It is the weight matrix of the bounding box regression branch. This is the corresponding bias vector. The Sigmoid function restricts the predicted coordinates to between 0 and 1, ensuring that the output is normalized coordinates.

[0164] IV. Specific Steps of S4

[0165] After the imaging-free target detection network outputs the predicted bounding box and target confidence score, the detection output module determines whether a target exists in the current observation area based on the target confidence score.

[0166] In one embodiment of the present invention, the specific steps for the detection output module to determine whether a target exists in the current observation area are as follows:

[0167] (1) If the target confidence level Greater than or equal to the preset confidence threshold (For example If the target is identified, the predicted bounding box will be output.

[0168] like If the target is not found, then it is determined that no target exists. The predicted bounding box consists of normalized coordinates relative to the training image size, which can be restored to pixel coordinates by multiplying by the training image size.

[0169] The significance of this decision mechanism is that when the model determines that there is no target in the sample, the output of the bounding box regression branch does not participate in subsequent processing, thus avoiding meaningless localization results from samples without targets; when the model determines that there is a target in the sample, the bounding box localization result is output simultaneously, realizing the one-step completion of target existence judgment and spatial location prediction.

[0170] (2) If If no target is found, a background determination signal for the current observation area is output. This background determination signal triggers the detection process for the next observation area of ​​the scene to be detected and reduces the priority of subsequent processing and transmission for the current observation area. This background determination signal mechanism transforms the "no target" determination result from the algorithm's internal state into a system-level control signal, enabling automatic advancement of the detection process and hierarchical management of resources, effectively reducing the consumption of storage, bandwidth, and computing resources by targetless areas.

[0171] To verify the effectiveness of the imaging-free target detection method based on single-pixel measurement provided by this invention, specific experimental verification was conducted using an embodiment of a rapid detection scenario for maritime ship targets.

[0172] (1) Experimental setup

[0173] This experiment conducted general performance verification experiments and application scenario-specific verification experiments on publicly available detection datasets.

[0174] The specific experimental setup is as follows:

[0175] The training images are all 64×64 pixels in size;

[0176] The sampling rate was set to 25%, and the number of single-pixel measurements was M=1024; the dimensions of the fully connected layers in the multilayer perceptron feature extraction module were 1024→512→256→128→64 in sequence; the confidence threshold T=0.5; and the IoU threshold was set to 0.4 (when the intersection-union ratio of the predicted box and the ground truth box is greater than or equal to 0.4, the target is considered to have been successfully located).

[0177] The coordinate loss weight λcoord = 10.0, and the targetless loss weight λnoobj = 0.8. A two-stage training strategy was adopted: the first stage had a learning rate of 0.001 and was trained for 150 epochs; the second stage had a learning rate of 0.0001 and was trained for 100 epochs. The experiment used Ubuntu 20.04 operating system, PyTorch 1.12 deep learning framework, and the hardware configuration was Intel Xeon Gold 6248 CPU, NVIDIA RTX 3090 GPU (24GB VRAM). The batch size was set to 32, and the optimizer used was Adam.

[0178] Evaluation indicators include:

[0179] Detection accuracy P = TP / (TP + FP) is used to measure the model's ability to control false alarms;

[0180] Detection recall R = TP / (TP + FN) is used to measure the model's ability to detect targets;

[0181] The positioning success rate, Areg = NIoU ≥ 0.4TP, is used to evaluate the model's ability to accurately predict the target's location after successfully determining its existence.

[0182] Here, TP represents true positives, which are the number of samples that are actually containing a target and are correctly identified as containing a target by the model; FP represents false positives, which are the number of samples that are actually not containing a target but are incorrectly identified as containing a target by the model; FN represents false negatives, which are the number of samples that are actually containing a target but are incorrectly identified as not containing a target by the model; NIoU≥0.4 means that among the successfully detected samples containing a target, the number of samples whose predicted bounding box and bounding box label have an intersection-union ratio greater than or equal to 0.4.

[0183] (2) Analysis of experimental results

[0184] Experimental results for target existence determination: The imaging-free target detection network of this invention achieved a detection precision of 84.28% and a detection recall of 90.59% on the test set. The high recall (90.59%) indicates that the model can detect most real targets, demonstrating good target detection capability. This verifies that effective information related to target existence is still retained in the low-dimensional measurement vector, and the MLP feature extraction module can effectively extract this information. The high precision (84.28%) indicates that the model has good background suppression capability and can distinguish target samples from background samples to a certain extent.

[0185] The results show that after two-stage training, the measurement pattern learned by the single-pixel modulation module enables the target sample and the non-target sample to form a distinguishable response distribution in the measurement domain. The MLP feature extraction module can extract discriminative features from this measurement distribution, and the confidence branch can make an accurate binary classification judgment based on these features.

[0186] Experimental results for target location prediction: The model achieved a 54% localization success rate on the test set. For samples with clear target outlines, weak background interference, and moderate target size, the predicted bounding boxes maintained good overlap with the ground truth bounding boxes, indicating that the low-dimensional measurement vectors still contain some spatial location information. The MLP network and bounding box regression branch possess target localization capabilities to some extent. However, the localization success rate was lower than the detection precision and recall. This is because: single-pixel modulation compresses two-dimensional spatial information into a one-dimensional measurement array, and the edges and outlines of the target are no longer preserved in an intuitive form. The network needs to infer the target coordinates from the implicit relationships between the measurement values, which is more difficult to learn; the MLP structure itself lacks the local receptive field and translation sharing mechanism found in convolutional neural networks, which limits its ability to perform fine-grained localization.

[0187] Bounding box regression is more difficult than confidence prediction because target location information is distributed more implicitly in the measurements during single-pixel modulation, requiring the network to learn more complex mapping relationships.

[0188] (3) Comparison experiment of single-pixel modulation matrix and comparison experiment of different sampling rates

[0189] To verify the effectiveness of the trainable single-pixel modulation matrix, this embodiment conducts comparative experiments using both a randomly fixed single-pixel modulation matrix and a trainable single-pixel modulation matrix. Please refer to the appendix. Figure 6 , Figure 6 This is a comparison chart of experimental results for different modulation matrices used in the experiments of this invention. Figure 6 The comparison of detection precision, detection recall, and localization success rate between random fixed modulation matrices and trainable modulation matrices is presented. Random fixed modulation matrices are randomly initialized (uniformly distributed 0-1) before training and their parameters are not updated during training; trainable modulation matrices, on the other hand, serve as learnable parameters for the network front-end and participate in training together with the back-end detection network.

[0190] The trainable modulation matrix outperforms the random fixed modulation matrix in terms of detection precision, recall, and localization success rate. Specifically, compared to the random fixed modulation matrix, the trainable modulation matrix improves detection precision by approximately 4.09% (80.19%→84.28%), detection recall by approximately 5.39% (85.20%→90.59%), and localization success rate by approximately 13.49% (40.51%→54.00%). The improvement in localization success rate is the largest, indicating that the trainable modulation matrix not only helps determine the presence of a target but also better preserves the target's spatial location information. This result verifies that when the single-pixel modulation matrix is ​​optimized end-to-end based on backpropagation of the detection loss, the measurement process can adaptively prioritize preserving information related to target detection rather than information related to image reconstruction, thereby improving the information utilization efficiency of the measurement data.

[0191] Please see the appendix Figure 7 , Figure 7 The figure shows the performance variation curves of the model under different sampling rates in the experiments of this invention. As can be seen from the figure, a comparative experiment was then conducted for different sampling rates, and the performance variation curves of the model under different sampling rates are shown below. Figure 7 As shown, comparative experiments were conducted with sampling rates of 12.5%, 25%, and 50% respectively. Figure 7The results show the trends of detection precision, detection recall, and localization success rate at these three sampling rates, corresponding to 512, 1024, and 2048 single-pixel measurements, respectively. The network structure, training strategy, and evaluation metrics were kept consistent throughout the experiment; only the dimension of the output measurement vector of the single-pixel modulation module was changed, and the input layer dimension of the MLP feature extraction module was adjusted accordingly. The experimental results are shown in Table 1.

[0192] Please refer to Table 1, which shows the performance comparison results of the model under different sampling rates in the embodiments of the present invention. The data in Table 1 shows that when the sampling rate is 12.5%, the number of measurements per pixel is relatively low (512), and the spatial information contained in the measurement vector is limited. Although the model can still complete the target existence judgment to a certain extent, with a single detection precision of 78.46% and a detection recall of 84.72%, the target localization ability is significantly affected, resulting in a localization success rate of only 41.35%. This indicates that under excessively low sampling rates, the spatial location information required for bounding box regression is insufficient, making it difficult for the bounding box regression task to converge.

[0193] Table 1. Comparison of model performance at different sampling rates

[0194]

[0195] When the sampling rate was increased to 25%, all three metrics showed significant improvement (detection precision 84.28%, detection recall 90.59%, and localization success rate 54.00%). This indicates that at this sampling rate, the single-pixel measurement vector can effectively preserve target presence information and some spatial location information, meeting the basic requirements of imaging-free target detection tasks. Furthermore, compared to a 50% sampling rate, a 25% sampling rate requires fewer measurements, reducing data acquisition and computational overhead while maintaining detection performance.

[0196] When the sampling rate is further increased to 50%, theoretically there is more sufficient measurement information, which is beneficial for preserving spatial features in the target scene. However, in practice, the performance not only did not continue to improve, but also declined to some extent (detection precision decreased by about 2.11%, detection recall decreased by about 2.16%, and localization success rate decreased by about 4.74%). The reason for this is that increasing the sampling rate means increasing the dimension of the input measurement vector (from 1024 dimensions to 2048 dimensions), thus increasing the amount of information and feature complexity that the model needs to process. Under the condition of a limited training sample size, an excessively high sampling rate causes the model to rely excessively on local features and redundant measurement information in the training set. This enhances the model's ability to fit the training data but weakens its generalization ability, leading to overfitting. Furthermore, the model's failure to converge sufficiently under the same number of training epochs is also a possible reason for the performance decline.

[0197] Based on the experimental results of the three sampling rates, a 25% sampling rate achieved the best balance between measurement information, model generalization ability, and data processing cost. This experimental conclusion verifies the rationality of setting the sampling rate between 12.5% ​​and 50%, with 25% being the preferred value, in this invention. This numerical range cannot be obtained through simple experiments; a higher sampling rate is not always better, as excessively high sampling rates can lead to overfitting. There exists an optimal range and a corresponding best balance point.

[0198] (4) Scenario Experiment Verification

[0199] Specifically, in the scenario of detecting ships at sea, the method was experimentally validated based on a publicly available dataset. The experimental setup was consistent with the dataset.

[0200] Experimental results show that the model achieved a detection precision of 90.36%, a detection recall of 88.56%, and a localization success rate of 45.79% on the public test set.

[0201] Compared to the experimental results on the public dataset, the detection precision improved (from 84.28% to 90.36%). This is attributed to the differences in texture and grayscale distribution between ship targets and the sea background in the public dataset. After single-pixel modulation, the target samples and background samples exhibit a more significant difference in response in the measurement domain, which is beneficial for the confidence prediction branch to distinguish them. The detection recall decreased slightly (from 90.59% to 88.56%). This is because some ship targets in the public dataset are small and have low contrast with the sea background, resulting in some loss of target feature information during compressed measurement, leading to a small number of missed detections. However, the model still maintained high detection precision and recall, further validating the effectiveness of this invention under different data distributions and target feature conditions.

[0202] Furthermore, the localization success rate decreased slightly (from 54% to 45.79%). This is because some ship targets in the public dataset are relatively small, occupying only a few pixels in the 64×64 training images. After single-pixel modulation at a sampling rate of 25%, the proportion of target location information in the measurement vector is further reduced, increasing the difficulty of bounding box regression. This result also indicates that for application scenarios with a high proportion of small targets, the sampling rate can be appropriately increased to retain more spatial location information.

[0203] (5) Analysis of reasoning speed

[0204] To verify the applicability of the method of this invention on resource-constrained platforms, the model inference speed was tested on an edge computing device (NVIDIA Jetson Xavier NX, 15W power mode). The model parameters are approximately 8.7M, of which the single-pixel modulation matrix accounts for 4.2M, the MLP feature extraction module accounts for 3.9M, and the dual-branch detection head accounts for 0.6M. The measurement vector generation time for a single inference depends on the hardware refresh rate of the spatial light modulator; the total acquisition time for 1024 measurements is approximately 3.1s; the network forward inference time is approximately 2.3ms. Measurement acquisition and inference can be pipelined in parallel, with a total processing time of approximately 3.1s for a single observation.

[0205] In scenarios requiring only rapid initial screening (e.g., quickly determining the presence of a target in a certain area), a lower sampling rate (e.g., 12.5%) can be used. In this case, the measurement and acquisition time is shortened to approximately 1.6 seconds, at the cost of a decrease in the positioning success rate of approximately 16.11% (from 54.00% to 41.35%). Users can flexibly choose the sampling rate based on the accuracy and timeliness requirements of their actual application scenarios.

[0206] During the testing phase, the model processed 1849 images in a total time of approximately 3.29 seconds, with an average processing time of about 1.78 ms per image and an average processing speed of approximately 561 images per second. This speed advantage mainly stems from the use of low-dimensional measurement data (selecting a 25% sampling rate) as input and the model structure being primarily based on an MLP, which is simpler compared to complex image detection network structures. Combined with the high-speed response characteristics of a single-pixel detector, this invention possesses excellent real-time detection potential.

[0207] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for imaging-free target detection based on single-pixel measurement, characterized in that, Includes the following steps: S1: Construct a sample set to train an imaging-free target detection network and a single-pixel modulation module; S2: The single-pixel detection system acquires a one-dimensional measurement array of the current observation area; S3: Transmit the one-dimensional measurement array to the trained imaging-free target detection network to obtain the target confidence and predicted bounding box of the one-dimensional measurement array; S4: The detection output module determines whether there is a target in the current observation area based on the target confidence level. If there is a target, the target's location is marked according to the predicted bounding box. If there is no target, the current observation area is changed, and the process proceeds to step S2. The imaging-free target detection network includes a multilayer perceptron feature extraction module and a dual-branch detection module. Step S3 specifically includes: S31: The multilayer perceptron feature extraction module performs nonlinear feature mapping on the one-dimensional measurement array to obtain the target feature vector; S32: The dual-branch detection module extracts features from the target feature vector and simultaneously outputs the target confidence score and the predicted bounding box.

2. The imaging-free target detection method according to claim 1, characterized in that, In step S2, the single-pixel detection system uses the trained single-pixel modulation matrix of the single-pixel modulation module to encode the spatial information of the current observation area.

3. The imaging-free target detection method according to claim 1, characterized in that, The specific steps of S31 include: the multilayer perceptron feature extraction module uses several fully connected layers linked end-to-end to compress the one-dimensional measurement array layer by layer; in each fully connected layer, the following steps are performed: The fully connected layer performs an affine mapping on the input vector to extract a low-dimensional vector. The activation layer performs a non-linear transformation on the low-dimensional vector to extract feature vectors; the activation layer is activated using the ReLU function. During training, the Dropout layer randomly sets some elements of the feature vector to zero and outputs them as an output vector to the next fully connected module. During testing, the Dropout layer directly uses the feature vector as the output vector and outputs it to the next fully connected module. The dimension of the input vector of each of the fully connected modules is greater than the dimension of the output vector; The random drop rate of the Dropout layer is set to 0.

5.

4. The imaging-free target detection method according to claim 3, characterized in that, The plurality of fully connected modules include a first-layer fully connected module, a second-layer fully connected module, a third-layer fully connected module, and a fourth-layer fully connected module; The dimension of the one-dimensional measurement array is set to 1024; The dimension of the output vector of the first fully connected module is set to 512; The dimension of the output vector of the second fully connected module is set to 256; The dimension of the output vector of the third fully connected module is set to 128; The dimension of the output vector of the fourth fully connected module is set to 64.

5. The imaging-free target detection method according to claim 1, characterized in that, The specific steps of S32 include: the dual-branch detection module extracts the target confidence from the target feature vector through the confidence prediction branch, and extracts the predicted bounding box from the target feature vector through the bounding box regression branch; The calculation process for extracting the target confidence level is as follows: ， The calculation process for extracting the predicted bounding box is as follows: ， in, The target confidence level, Let H be the predicted bounding box, Sigmoid be the activation function, h be the target feature vector, and W be the target bounding box. conf Let b be the weight matrix of the confidence prediction branch. conf W is the bias vector of the confidence prediction branch. reg Let b be the weight matrix of the bounding box regression branch. reg The bias vector of the bounding box regression branch; The target confidence level is set as a one-dimensional vector; the predicted bounding box is set as a four-dimensional vector.

6. The imaging-free target detection method according to claim 5, characterized in that, The specific steps of S32 also include performing independent dimensionality adjustment processing on the target feature vector in the confidence prediction branch and the bounding box regression branch, respectively.

7. The imaging-free target detection method according to claim 1, characterized in that, The specific steps of S1 include: S11: Construct a sample set, which includes target samples and non-target samples; the sample structure of the sample set includes training images, confidence labels, and bounding box labels; S12: The training image is converted into the one-dimensional measurement array by the single-pixel modulation module; S13: Input the one-dimensional measurement array into the imaging-free target detection network to obtain the target confidence and the predicted bounding box; S14: Calculate the loss function based on the target confidence and the predicted bounding box. If the loss function meets the preset conditions, end the training. If the loss function does not meet the preset conditions, jointly adjust the parameters of the single-pixel modulation module and the imaging-free target detection network, and return to S12. S15: Obtain the trained single-pixel modulation module and the trained imaging-free target detection network.

8. The imaging-free target detection method according to claim 7, characterized in that, The specific steps of S12 include constructing the single-pixel modulation module to simulate the spatial coding and compressed sampling process of the single-pixel detection system, specifically: The image conversion module unfolds the training image, converts it into one-dimensional data in equal quantities, and performs pixel normalization processing. A single-pixel modulation matrix is ​​used to compress and sample normalized one-dimensional data to obtain the original one-dimensional measurement array. The original one-dimensional measurement array is batch normalized and nonlinearly activated to output the one-dimensional measurement array.

9. The imaging-free target detection method according to claim 7, characterized in that, The specific steps in S14 include: (1) Calculate the confidence loss : ； in, The target confidence level, The confidence level label; (2) Calculate the bounding box regression loss ; ； in, The intersection-union ratio (IoU) of the predicted bounding box and the bounding box label is given. The square of the Euclidean distance between the predicted bounding box and the center point of the bounding box label; The square of the diagonal length of the minimum bounding rectangle of the predicted bounding box and the bounding box label; This is a consistent aspect ratio item; These are the weighting coefficients; (3) Calculate the loss of the target sample. : ； in, Weights for coordinate loss; (4) Calculate the loss of the target sample. : ； in, Weights for loss without a target; (5) Calculate the loss function L for training in the same batch: ； in, Indicates batch size; Indicates the first sample in the sample set The total loss for the nth sample; when the nth sample... When each sample is the target sample. When the first When the sample is the target sample, .

10. A single-pixel measurement-based target detection system without imaging, employing the method described in claim 1 to detect targets in a scene, characterized in that... include: A single-pixel detection system is used to acquire a one-dimensional measurement array of the current observation area; An imaging-free target detection network includes a multilayer perceptron feature extraction module and a dual-branch detection module. The multilayer perceptron feature extraction module is used to perform nonlinear feature mapping on the one-dimensional measurement array to obtain the target feature vector. The dual-branch detection module is used to simultaneously calculate the target confidence and predict the bounding box based on the target feature vector. The detection output module is used to identify the target in the current observation area based on the target confidence level and the predicted bounding box.

Images