Single-pixel initial reconstruction image enhancement method, system, medium and identification method

Abstract

The present application relates to a kind of single-pixel initial reconstruction image enhancement method, system, medium and identification method in the field of computational imaging technology.Single-pixel image reconstruction method includes: first, the initial reconstruction image Y of the single-pixel imaging to be reconstructed is carried out multi-scale feature extraction, and global degradation condition vector c and spatial reliability graph R are generated.Next, Y, c and R are carried out by coarse to fine progressive reconstruction, and coarse reconstruction result Xc and detail residual information ΔX are obtained.Subsequently, according to Y, R, Xc and ΔX, generate gate value G.And G and R generate fusion weight W.Finally, based on W, Xc and ΔX are weighted fusion, and final reconstruction image is obtained.The present application explicitly models sampling rate and spatial degradation features through physical condition encoder to realize degradation perception.Through spatial reliability graph, non-uniform degradation is processed specifically, and artifact and structural distortion are suppressed.Through progressive reconstruction, coarse to fine step recovery is realized, and structure fidelity and detail richness are balanced.

Description

Technical Field

[0001] This invention relates to the field of computational imaging technology, and in particular to a method, system, medium, and recognition method for enhancing a single-pixel initial reconstructed image. Background Technology

[0002] Single-Pixel Imaging (SPI) is a novel computational imaging technique that utilizes a single-pixel detector instead of a traditional array detector to acquire image information. SPI modulates the target scene through a series of spatial light modulation modes, then the single-pixel detector records the corresponding light intensity values, and finally, a reconstruction algorithm reconstructs the target image. Compared with traditional area array imaging, SPI has unique advantages in non-visible light band imaging, low-light imaging, and scattering medium imaging. However, the imaging quality of SPI is highly dependent on the sampling rate. In practical applications, to shorten the acquisition time, low sampling rates (such as 1%~25%) are usually used for sampling. This leads to severe degradation of the initial reconstructed image (usually obtained by traditional methods such as Compressed Sensing (CS) and Filtered Back Projection (FBP)). The main manifestations are: a large amount of missing structural information, obvious fringe artifacts, local holes and breaks, and loss of high-frequency details. These degradations are not simple additive noise, but complex degradations with global and structural characteristics. In existing technologies, methods for reconstructing initial images mainly include the following categories: CNN-based image restoration methods (such as DnCNN and UNet), Transformer-based image restoration methods (such as Uformer, SwinIR, and Restormer), and Generative Adversarial Network (GAN)-based image enhancement methods. However, the fixed restoration strategies employed in these reconstruction methods are difficult to adapt to different sampling rates such as 2%, 5%, and 25%, and applying the same enhancement intensity to all spatial locations may amplify artifacts in low-confidence regions. Furthermore, one-step prediction may introduce erroneous high-frequency details before the main structure has stabilized. Summary of the Invention

[0003] To address the technical problem that existing SPI image reconstruction methods are prone to structural distortion and artifacts at low sampling rates, this invention provides a single-pixel initial reconstruction image enhancement method, system, medium, and recognition method.

[0004] In a first aspect, the present invention provides a physical condition-guided single-pixel initial reconstruction image enhancement method, comprising: Obtain the initial reconstructed image Y of the single-pixel imaging to be enhanced and its corresponding sampling rate α. Calculate the gradient magnitude map Gm and Laplacian response map L of Y using a physical conditional encoder, and expand the sampling rate α into a sampling rate feature map A with the same spatial dimensions as Y. α And map the sampling rate α to the sampling rate embedding vector e. Let Y, Gm, L, and A... α The physical augmentation input P is formed by concatenating along the channel dimension. Multi-scale feature encoding is performed on P to obtain multi-scale features. The deepest feature in the multi-scale features is combined with e to generate a global degradation condition vector c. A spatial reliability map R is generated based on the multi-scale features and c.

[0005] Y is input into a coarse reconstruction image restoration network, and c and R are used to modulate the network to obtain the coarse reconstruction result Xc. Y, Xc, and R are concatenated and input into a detail residual image restoration network, and c and R are used to modulate the network to obtain the detail residual ΔX. A gate value G is generated based on the residual energy of Y, Xc, R, and ΔX, and a fusion weight W is generated based on G and R. Using Xc as a base, the W-gated ΔX is added to Xc and constrained to a preset image value range to obtain the final enhanced image X.

[0006] Secondly, this invention proposes a physical condition-guided single-pixel initial reconstruction image enhancement system, which uses the physical condition-guided single-pixel initial reconstruction image enhancement method from the first aspect. The single-pixel initial reconstruction image enhancement system includes: an input module, a physical condition encoding module, a progressive reconstruction module, and a reliability fusion module.

[0007] The input module acquires the initial reconstructed image Y of the single-pixel imaging to be enhanced and its corresponding sampling rate α. The physical conditional encoding module calculates the gradient magnitude map Gm and Laplacian response map L of Y using a physical conditional encoder, and expands the sampling rate α into a sampling rate feature map A with the same spatial size as Y. α And map the sampling rate α to the sampling rate embedding vector e. Let Y, Gm, L, and A... αThe physical enhancement input P is formed by concatenating along the channel dimension. Multi-scale feature encoding is performed on P to obtain multi-scale features. The deepest feature in the multi-scale features is combined with e to generate a global degradation conditional vector c, and a spatial reliability map R is generated based on the multi-scale features and c. The progressive reconstruction module inputs Y into the coarse reconstruction image restoration network and modulates the coarse reconstruction image restoration network using c and R to obtain the coarse reconstruction result Xc. Y, Xc, and R are concatenated and input into the detail residual image restoration network, and the detail residual image restoration network is modulated using c and R to obtain the detail residual ΔX. The reliability fusion module generates a gate value G based on the residual energy of Y, Xc, R, and ΔX, and generates fusion weights W based on G and R. Based on Xc, the W-gated ΔX is added to Xc and constrained to a preset image value range to obtain the final enhanced image X.

[0008] Thirdly, this invention proposes an aerial target identification method, which includes the following steps: Acquire single-pixel imaging measurement data and corresponding sampling rates in the target airspace. Form an initial reconstructed image Y using compressed sensing reconstruction or filtered back projection. Enhance the initial reconstructed image Y using the physical condition-guided single-pixel initial reconstructed image enhancement method described in the first aspect to obtain the final enhanced image X. Input the final enhanced image X into the target recognition model to obtain the category and / or location of the aerial target.

[0009] Fourthly, the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon. When executed by a processor, the computer program / instructions implement the steps of the physically condition-guided single-pixel initial reconstruction image enhancement method of the first aspect.

[0010] The beneficial effects of this invention are as follows: 1. This invention explicitly models the sampling rate and spatial degradation features using a physical condition encoder, adaptively adjusting the recovery strategy to significantly improve PSNR (Peak Signal-to-Noise Ratio) at extremely low sampling rates, thus achieving degradation perception. It specifically addresses uneven degradation through spatial reliability maps, suppressing artifacts and structural distortion to achieve spatial adaptive processing. Progressive reconstruction achieves coarse-to-fine step-by-step recovery, balancing structural fidelity and detail richness. Through multi-dimensional constraints, it utilizes spatial and frequency domain losses to synergistically suppress artifacts, effectively suppressing SPI-specific stripe artifacts and improving visual quality.

[0011] 2. Compared with image restoration methods that do not incorporate physical conditions and spatial reliability, this invention can employ different conditional modulations for different sampling rates; it can distinguish the restoration credibility of target areas, edge areas, and artifact areas; and it can separate the restoration of the main structure from the compensation of local details. Qualitative reconstruction results show improved continuity of the target's outer contour and slender structures, clearer edge grayscale transitions, and enhanced local texture and brightness levels, while mitigating stripe artifacts, block artifacts, local breaks, holes, and oversmoothing. This improvement does not rely on generative adversarial loss, thus helping to preserve the true shape of targets such as aircraft, birds, and drones.

[0012] 3. This invention fully combines the advantages of SPI physical priors and deep learning, achieving a dual improvement in structural fidelity and detail richness in low sampling rate SPI reconstruction tasks, which is significantly better than existing technical solutions. Attached Figure Description

[0013] 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 these drawings without creative effort.

[0014] Figure 1 This is a flowchart of a physical condition-guided single-pixel initial reconstruction image enhancement method; Figure 2 This is a comparison diagram of the visual information of the reconstructed images from existing methods and the method of this invention. Detailed Implementation

[0015] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0016] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0017] This invention provides a physical condition-guided single-pixel initial reconstruction image enhancement method. It achieves high-quality, high-fidelity SPI image reconstruction at low sampling rates by explicitly encoding SPI physical degradation conditions, predicting spatial reliability maps, and employing a coarse-to-fine progressive reconstruction strategy. For details, please refer to... Figure 1 The single-pixel image reconstruction method includes the following steps: First, obtain the initial reconstructed image Y of the single-pixel imaging to be enhanced and its corresponding sampling rate α. Then, calculate the gradient magnitude map Gm and Laplacian response map L of the initial reconstructed image Y using a physical conditional encoder, and expand the sampling rate α into a sampling rate feature map A with the same spatial size as the initial reconstructed image Y. α The sampling rate α is mapped to the sampling rate embedding vector e; the initial reconstructed image Y, gradient magnitude map Gm, Laplacian response map L, and sampling rate feature map A are then mapped. α The physical augmentation input P is formed by concatenating along the channel dimension. Multi-scale feature encoding is then performed on the physical augmentation input P to obtain multi-scale features. The deepest feature in the multi-scale features is combined with the sampling rate embedding vector e to generate a global degradation condition vector c. A spatial reliability map R is generated based on the multi-scale features and the global degradation condition vector c. The initial reconstructed image Y is then input into a coarse reconstruction image restoration network, and the global degradation condition vector c and the spatial reliability map R are used to modulate the coarse reconstruction image restoration network to obtain the coarse reconstruction result Xc. The initial reconstructed image Y, the coarse reconstruction result Xc, and the spatial reliability map R are concatenated and input into a detail residual image restoration network, and the global degradation condition vector c and the spatial reliability map R are used to modulate the detail residual image restoration network to obtain the detail residual ΔX. Finally, a gate value G is generated based on the residual energy of the initial reconstructed image Y, the coarse reconstruction result Xc, the spatial reliability map R, and the detail residual ΔX. Fusion weights W are generated based on the gate value G and the spatial reliability map R. Based on the coarse reconstruction result Xc, the detail residual ΔX after fusion weight W gate is added to the coarse reconstruction result Xc and restricted to the preset image value range to obtain the final enhanced image X.

[0018] I. Physical Condition Coding.

[0019] One of the core innovations of this invention is physical conditional encoding, which can explicitly model sampling rate and spatial degradation features. Specifically, the input of the physical conditional encoder includes a single-channel initial reconstructed image Y and a sampling rate α. The initial reconstructed image Y has a size of 1×H (height)×W (width), and the pixel values ​​are normalized to [0,1]. The sampling rate α can be input as a proportional value, for example, 2%, 5%, and 25% correspond to 0.02, 0.05, and 0.25, respectively. The sampling rate encoding branch generates a sampling rate embedding vector e through two layers of fully connected mapping and GELU activation, while broadcasting the sampling rate α as a sampling rate feature map A with the same spatial size as the initial reconstructed image Y. αIn the physical condition encoder, the Sobel operator is used to obtain the horizontal and vertical gradients, and the Laplacian operator is used to obtain the second-order response, thus obtaining the corresponding gradient magnitude map Gm and Laplacian response map L, which are represented as follows:

[0020]

[0021] Among them, Mean c (·) indicates that the arithmetic mean is calculated along the channel dimension of the input tensor. S represents the convolution operation. x and S y Represents the Sobel core, K L Let represent the Laplace kernel, and ε represent a positive number to prevent numerical instability. The physics-enhanced input P is represented as:

[0022] The physical augmentation input P is processed through an initial convolutional layer and a three-stage stride convolutional layer for multi-scale feature encoding, resulting in multi-scale features, namely, features F0, F1, F2, and F3 at four scales in this embodiment. The deepest feature F3 is then concatenated with the sampling rate embedding vector e after global average pooling, and finally mapped through two fully connected layers to generate a global degradation conditional vector c.

[0023] GAP(·) represents global average pooling. c (·) denotes the global conditional mapping function. Simultaneously, features at each scale are upsampled to H×W and concatenated along the channel dimension. After dimensionality reduction, they are concatenated with the conditional feature map obtained from the global degenerate conditional vector c. Finally, a spatial reliability map R is generated through convolution and the sigmoid function.

[0024] Where σ(·) represents the Sigmoid function. R (·) represents the convolution mapping function, Up(·) represents upsampling, and Broadcast(·) represents the broadcast mechanism. The spatial reliability map R has a size of 1×H×W and a value range of [0,1], representing the relative reconstruction reliability of each spatial location. The sampling rate embedding vector e is not directly concatenated with the two-dimensional image features. The sampling rate feature map A participates in the two-dimensional concatenation. α The sampling rate embedding vector e is then used to generate the global condition vector c.

[0025] Furthermore, the architecture of this physical conditional encoder includes: a sampling rate coding branch, a physical prior extraction branch, a multi-scale feature coding branch, and a global conditional generation branch. The sampling rate coding branch includes: two fully connected layers and a GELU activation layer sandwiched between them, used to expand the sampling rate α into a sampling rate feature map A with the same spatial size as the initial reconstructed image Y. α The sampling rate α is mapped to a sampling rate embedding vector e. The physical prior extraction branch includes: a Sobel gradient operator with a fixed kernel, a Laplacian operator, and a stitching layer. The Sobel gradient operator and the Laplacian operator are used to process the initial reconstructed image Y to obtain the corresponding gradient magnitude map Gm and Laplacian response map L, respectively. The stitching layer is used to process the initial reconstructed image Y, the gradient magnitude map Gm, the Laplacian response map L, and the sampling rate feature map A. α The input P is formed by concatenating along the channel dimension. The multi-scale feature encoding branch includes an initial convolutional layer and three stride convolutional layers connected sequentially. The kernel size of each convolutional layer is 3×3. Each convolutional layer is followed by a GELU activation layer. This layer is used to downsample the physical augmentation input P to one-eighth of its original size, obtaining feature maps at four scales, namely the features F0, F1, F2, and F3 mentioned above. The global condition generation branch includes an adaptive global average pooling layer, two fully connected layers, and a GELU activation layer located between the two fully connected layers. The adaptive global average pooling layer concatenates the deepest feature F3 in the multi-scale features with the sampling rate embedding vector e after global average pooling, and obtains the global degradation condition vector c through fully connected mapping. The spatial reliability generation branch includes an upsampling layer, a 1×1 convolutional layer, two 3×3 convolutional layers, and a Sigmoid layer. It can upsample feature maps of four scales to the original input size and then concatenate them along the channel dimension. After dimensionality reduction by 1×1 convolution and GELU activation, it is concatenated along the channel with the conditional feature map obtained by expanding the global degenerate conditional vector c. Then, it is successively concatenated by 3×3 convolution, GELU activation, 3×3 convolution and Sigmoid normalization to generate a spatial reliability map R.

[0026] II. Gradual Reconstruction from Coarse to Fine.

[0027] After obtaining the global degradation condition vector c and spatial reliability map R through physical condition encoding, the system proceeds to the coarse reconstruction stage and the detail residual stage. In this embodiment, these two stages are implemented using two different specifications of image restoration networks. The image restoration network can adopt a Uformer network architecture, which includes: an input layer, multiple encoders, a bottleneck layer, multiple decoders, and an output layer. The encoder includes: a LeWinTransformer block and a downsampling layer. The bottleneck layer includes: multiple LeWinTransformer blocks. For ease of distinction, the image restoration network in the coarse reconstruction stage is named the coarse reconstruction image restoration network. The image restoration network in the detail residual stage is named the detail residual image restoration network. The difference between the two specifications is that the embedding dimension of the detail residual image restoration network is half that of the coarse reconstruction image restoration network. The number of network layers in the detail residual image restoration network is one-third that of the coarse reconstruction image restoration network, and it is used to compensate for local details based on the restored main structure. The coarse reconstruction image restoration network receives the initial reconstructed image Y, and receives the global degradation condition vector c and spatial reliability map R in each Transformer block, and outputs the coarse reconstruction result Xc. The detail residual image restoration network receives inputs formed along the channel dimension from the initial reconstructed image Y, the coarse reconstruction result Xc, and the spatial reliability map R. It also receives the global degradation condition vector c and the spatial reliability map R, and outputs the detail residual information ΔX. The correspondence is expressed as:

[0028]

[0029] F c (·) represents the nonlinear mapping achieved by the coarse image reconstruction network, F d (·) denotes the nonlinear mapping implemented by the detail residual image restoration network; c and R after the semicolon represent the global degradation condition vector and spatial reliability map, which serve as the inputs to the network modulation conditions. Specifically, in the coarse reconstruction stage, the initial reconstructed image Y is first converted into a token sequence with an embedding dimension of 32, and then sequentially enters the encoder, bottleneck layer, and decoder for processing, outputting the coarse reconstruction result Xc through the output layer. Next, in the detail residual stage, the initial reconstructed image Y, the coarse reconstruction result Xc, and the spatial reliability map R are concatenated and input into the detail residual image restoration network for processing to obtain the detail residual information ΔX.

[0030] III. Global Conditional Modulation and Reliability-Guided Window Attention Mechanism.

[0031] It is worth mentioning that a major core improvement of this invention is that FiLM-style conditional modulation and a reliability-guided window attention mechanism are applied to each layer of the LeWinTransformer block in the coarse reconstruction image restoration network and the detail residual image restoration network.

[0032] In FiLM-style conditional modulation, the global degradation conditional vector c generates scaling parameter γ(c) and offset parameter β(c) through two independent linear mappings, and FiLM-style modulation is performed on the normalized features of the attention sublayer and the feedforward network sublayer:

[0033]

[0034] In the formula, F′ represents the modulated feature, and F represents the normalized feature of the LayerNorm layer in the LeWinTransformer block. ⊙ denotes element-wise multiplication. tanh(·) represents the tanh function. f γ (·) and f β (·) denote the linear mappings used to generate the scaling parameter γ(c) and the offset parameter β(c), respectively. The scaling parameter γ(c) and the offset parameter β(c) are broadcast along the token or spatial dimension. Modulation after normalization at the LayerNorm layer ensures modulation stability. The scaling term is limited to the range [-0.1, 0.1], and the offset term is generated by independent linear mappings.

[0035] The reliability-guided window attention mechanism includes: The mechanism extracts spatial structure features from the input features to obtain an initial reliability score: Unlike ordinary window self-attention, which treats all locations equally, this mechanism first extracts the spatial structure of the input features channel-by-channel using depthwise separable convolution, obtaining an initial reliability log-odds z for each location and attention head. Here, the initial reliability log-odds z is a real-valued reliability score before Sigmoid normalization. Subsequently, the global degradation conditional vector c is scaled and biased to comprehensively correct the initial reliability score, allowing it to adaptively adjust according to the global degradation conditions. Then, the spatial reliability map R output by the physical conditional encoder is transformed into a log-odds prior with the same scale as the initial reliability score using a log-odds transformation, and this prior is superimposed on the initial reliability score (the ingenious aspect of log-odds is that it transforms probabilistic confidence into an additive score correction, directly integrating with the network's self-judgment score, achieving a unification of external physical prior and internal learning), ultimately yielding a reliability estimate. The reliability estimate is then used in two ways: First, it is used for position-wise weighted modulation of the reliability estimate and the input features (the reliability estimate is averaged head-to-head and multiplied by an amplification factor, which enhances the token features in high-reliability regions and suppresses low-reliability regions). Second, a pairwise attention bias constructed from the reliability estimate is added to the standard window self-attention calculation (the bias term is the product of the two positional reliability estimates minus the mean, making the attention weight between high-reliability regions higher), ultimately achieving "dual reliability guidance of feature content and attention connection," allowing the image restoration network to allocate more attention weights and feature enhancement intensity to high-reliability regions even with uneven SPI degradation, thereby suppressing artifacts and enhancing the real structure. Furthermore, the above reliability-guided window attention mechanism can be implemented by setting up a reliability-guided window attention module. This module includes: a depthwise separable convolutional layer, a conditional projection layer, a reliability estimation subnetwork, a pairwise bias generation subnetwork, and a weighted attention calculation subnetwork. The depthwise separable convolutional layer is used to extract the spatial structural features of the input features to obtain the initial reliability score. The conditional projection layer generates scale / bias parameters for each attention head based on the global degradation conditional vector c, thus correcting the initial reliability score. The reliability estimation subnetwork combines spatial structure features and the spatial reliability map R to generate reliability estimates. The pairwise bias generation subnetwork constructs pairwise attention biases based on the reliability estimates. The weighted attention computation subnetwork performs window self-attention computation on the reliability-weighted input features and incorporates the pairwise attention biases into the standard window self-attention computation. Specifically, the reliability-guided window attention first obtains the initial log-probability z of reliability per position and per attention head through depthwise separable convolution. The scale parameter s(c) and bias parameter b(c) generated from the global degradation conditional vector c adaptively modulate the initial reliability log-probability z.The spatial reliability map R is constrained to the numerically stable interval and then subjected to a log-odds transformation, which is then used as an external prior superposition.

[0036]

[0037] in, Let represent the fused reliability log-odds obtained by modulating the initial reliability log-odds z using global conditions and superimposing spatial reliability priors. s(c) and b(c) represent the scale modulation parameter and offset modulation parameter generated from the global degradation condition vector c, respectively. η p represents the learnable weights of spatial reliability priors. r represents the position-by-position, attention-by-attention reliability score obtained after Sigmoid normalization, with a value range of [0,1]. x represents the token feature of the input reliability guidance window attention module within the current window. Represents the token characteristics after reliability residual modulation, η g The learnable weights represent the reliability feature modulation. head (·) denotes the arithmetic mean along the attention head dimension. This modulation is a residual enhancement, rather than directly multiplying the input features by R. The pairwise biases constructed from r are further added to the window attention logit:

[0038] In the formula, A represents the window attention score matrix before Softmax normalization; QK T Let represent the product of the query matrix Q obtained by mapping the modulated token features x′ to the key matrix K. Here, the superscript T denotes the matrix transpose, and d represents the feature dimension of a single attention head. B pos This represents the relative position offset matrix within the window. η b Represents the learnable weights for pairwise reliability bias. rr T Mean(rr) represents the outer product matrix obtained by multiplying the reliability scores of each position within the same attention head pairwise. T This means averaging along the two positional dimensions of the outer product matrix and broadcasting it to the original matrix size to center the pairwise reliability bias.

[0039] IV. Reliability-guided residual fusion.

[0040] After obtaining the coarse reconstruction result Xc and the detailed residual information ΔX, the reliability fusion process begins. First, the absolute values ​​of the detailed residual information ΔX are taken along each channel and averaged to obtain the residual energy map E. Then, the residual energy map E, the initial reconstructed image Y, the coarse reconstruction result Xc, and the spatial reliability map R are concatenated and input into a gating network for processing. After Sigmoid normalization, the gating value G is obtained.

[0041] Next, based on the gate value G and the spatial reliability map R, the fusion weights W and the final reconstructed image X are generated:

[0042]

[0043] In the formula, clamp(·,0,1) is a range constraint function that restricts the result to between 0 and 1. λ represents the fusion coefficient between the spatial reliability map and the learned gating value. In one embodiment, λ=0.5. The coarse reconstruction result Xc is directly retained as the main structural baseline. The fusion weight W only controls the intensity of the addition of detail residuals, thereby reducing the possibility of overcompensation and spurious texture generation in low-reliability regions.

[0044] V. Loss Function and Training Method.

[0045] It should be noted that the models and networks used in this invention need to be used after training, and the training method adopts an end-to-end joint training mode. In this embodiment, the training method includes: First, configure the training environment: batch size 4, learning rate 0.0002, and 30 training epochs. Start training using the AdamW optimizer (weight decay 0.02) with a learning rate scheduler that performs 3 rounds of warm-up followed by cosine annealing. For each training batch, take paired samples from the dataset (sharp image as target, low-quality SPI image as input) and perform one complete forward propagation. Let the sharp target image be T, and the stopping gradient operation be denoted as sg(·). The coarse reconstruction loss L... c Final reconstruction loss L f and Gating detail loss L d They are respectively:

[0046]

[0047]

[0048] Edge consistency loss L e The frequency domain consistency loss L is calculated from the difference in gradient magnitude between the predicted image Gm(X) and the target image Gm(T). freq Calculated from the difference in magnitude of the two-dimensional real Fourier transforms (FFTs):

[0049]

[0050] The reliability weakly supervised objective D is constructed jointly by the input error and the coarse reconstruction error. Let the reliability weakly supervised objective D represent the channel mean weighted sum of the input error and the coarse reconstruction error, and normalize it according to the spatial maximum value of each image:

[0051] Reliability target R tar and reliability loss L r Represented as:

[0052] Total loss L total Represented as:

[0053] w c w f w d w e w freq w r These are the weights for the corresponding terms. In one embodiment, the six weights are set to 1.0, 1.0, 0.5, 0.1, 0.05, and 0.05 respectively. Each weight is a training hyperparameter selected based on the validation set performance and is not limited to the values ​​mentioned above. Subsequently, the parameters of the model in this invention are updated once through mixed precision backpropagation to achieve end-to-end joint training. After each round, PSNR / SSIM is calculated on the validation set, and the model weights with the best validation metrics are retained as the final model. This iterative training continues until the set number of rounds is completed to achieve the optimal performance of the model. Of course, other training strategies can also be used, such as pre-training the coarse reconstruction image restoration network and then freezing it, training only the physical conditional encoder and the detail residual image restoration network. This training strategy can accelerate convergence and quickly verify the model performance.

[0054] In summary, this invention explicitly models the sampling rate and spatial degradation features using a physical condition encoder, adaptively adjusting the recovery strategy to significantly improve PSNR (Peak Signal-to-Noise Ratio) at extremely low sampling rates, thus achieving degradation perception. It specifically addresses non-uniform degradation through spatial reliability maps, suppressing artifacts and structural distortion to achieve spatial adaptive processing. Progressive reconstruction achieves coarse-to-fine step-by-step recovery, balancing structural fidelity and detail richness. Multi-dimensional constraints utilize spatial and frequency domain losses to synergistically suppress artifacts, effectively suppressing SPI-specific stripe artifacts and improving visual quality. This invention fully combines the advantages of SPI physical priors and deep learning, achieving a dual improvement in structural fidelity and detail richness in low-sampling-rate SPI reconstruction tasks, significantly outperforming existing technologies.

[0055] VI. Data Organization and Experimental Examples.

[0056] The experimental images were uniformly converted to 256×256 pixel single-channel grayscale images. The data included three types of sharp scenes: airplanes, birds, and drones, with 4,501 images of airplanes, 4,461 images of birds, and 3,912 images of drones, totaling 12,874 different scenes. For each sharp image, two initial reconstruction inputs, CS and FBP, were generated, with three sampling rates of 2%, 5%, and 25%, resulting in six experimental conditions: CS2, CS5, CS25, FBP2, FBP5, and FBP25. Each experimental condition contained 12,874 pairs of input and sharp target samples. For each experimental condition, the training, validation, and test sets were divided at approximately 80%, 10%, and 10% ratios, corresponding to 10,300 training samples, 1,287 validation samples, and 1,287 test samples. The partitioning results were consistent across different reconstruction methods and ablation settings to ensure that the same scene was used for pairing comparisons.

[0057] The experiments were implemented using PyTorch, with a hardware platform consisting of two NVIDIA GeForce RTX 3090 graphics cards. Training was performed using the AdamW optimizer with an initial learning rate of 0.0002 and a weight decay of 0.02, for 30 training epochs, employing 3 rounds of warm-up and cosine annealing for learning rate scheduling. Individual training could utilize micro-batch size 4 and 2-step gradient accumulation to form an effective batch size of 8. Under other parallel configurations, the effective batch size, learning rate, and sample traversal method remained consistent. After each training epoch, PSNR and SSIM were calculated on the validation set, and weights were saved according to the validation metrics. Comparative experiments included initial reconstruction input, CycleGAN, ESRGAN, Uformer, Gate-Uformer, reliability-guided Gate-Uformer, and the complete model of this invention. Ablation experiments were conducted by disabling physical conditional encoding, reliability-guided attention, detail residual branch, fusion gating, edge consistency loss, frequency domain consistency loss, and reliability loss to analyze the impact of each component on the reconstruction results.

[0058] The reconstruction results show that the PSNR of the method of this invention is improved by approximately 16.2% and 6.3% respectively compared with existing methods (CycleGAN and ESRGAN). Furthermore, as... Figure 2 As shown, Figure 2The image on the left is reconstructed using existing methods, while the image on the right is reconstructed using the method of this invention. It can be seen that the method of this invention preserves more edge features in terms of visual information. In summary, the enhanced image of this invention achieves improvements in the following aspects: improved continuity of structures such as aircraft wings and fuselage outlines, bird wing edges, and UAV arms; clearer high-contrast edges and local grayscale transitions; enhanced hierarchy and discernible details in weakly textured areas; reduced stripes, block artifacts, local holes, and breaks in low-sampling-rate inputs; and better preservation of the target's shape and positional relationships compared to methods that directly generate high-frequency content.

[0059] In another embodiment, a physical condition-guided single-pixel initial reconstruction image enhancement system is proposed, which uses the physical condition-guided single-pixel initial reconstruction image enhancement method described in the above embodiments. The single-pixel initial reconstruction image enhancement system includes: an input module, a physical condition encoding module, a progressive reconstruction module, and a reliability fusion module. The input module is used to acquire the single-pixel initial reconstruction image Y to be enhanced and its corresponding sampling rate α. The physical condition encoding module is used to calculate the gradient magnitude map Gm and Laplacian response map L of Y using a physical condition encoder, and to expand the sampling rate α into a sampling rate feature map A consistent with the spatial size of Y. α And map the sampling rate α to the sampling rate embedding vector e. Let Y, Gm, L, and A... α The physical enhancement input P is formed by concatenating along the channel dimension. Multi-scale feature encoding is performed on P to obtain multi-scale features. The deepest feature in the multi-scale features is combined with e to generate a global degradation conditional vector c, and a spatial reliability map R is generated based on the multi-scale features and c. The progressive reconstruction module inputs Y into a coarse reconstruction image restoration network and modulates the coarse reconstruction image restoration network using c and R to obtain the coarse reconstruction result Xc. Y, Xc, and R are concatenated and input into a detail residual image restoration network, and the detail residual image restoration network is modulated using c and R to obtain the detail residual ΔX. The reliability fusion module generates a gate value G based on the residual energy of Y, Xc, R, and ΔX, and generates fusion weights W based on G and R. Based on Xc, the W-gated ΔX is added to Xc and restricted to a preset image value range to obtain the final enhanced image X.

[0060] In another embodiment, an airport airborne target recognition method is proposed, comprising the following steps: First, single-pixel imaging measurement data and corresponding sampling rates of the airport airspace are acquired using a single-pixel detector. An initial reconstructed image Y is formed using a compressed sensing reconstruction method or a filtered back projection method. Then, the initial reconstructed image Y is enhanced using the physical condition-guided single-pixel initial reconstructed image enhancement method described in the above embodiment to obtain a final enhanced image X. Finally, the final enhanced image X is input into a target recognition model to obtain the category, location, and confidence score of the aircraft or other airborne targets. This target recognition model can employ YOLOv8, R-CNN, Faster R-CNN, or other detection models, all of which are existing target recognition models and will not be elaborated upon here.

[0061] In another embodiment, a computer-readable storage medium is also proposed, which stores a computer program / instructions. When the computer program / instructions are executed by a processor, they implement the steps of the physically-guided single-pixel initial reconstruction image enhancement method described in the above embodiments. The computer-readable storage medium may include, but is not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0062] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0063] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for enhancing a single-pixel initial reconstructed image guided by physical conditions, characterized in that, It includes: Obtain the initial reconstructed image Y of the single-pixel imaging to be enhanced and its corresponding sampling rate α; The gradient magnitude map Gm and Laplace response map L of Y are calculated using a physical condition encoder, and the sampling rate α is expanded into a sampling rate feature map A with the same spatial size as Y. α And map the sampling rate α to the sampling rate embedding vector e; and map Y, Gm, L, and A α The physical augmentation input P is formed by concatenating along the channel dimension. Multi-scale feature encoding is performed on P to obtain multi-scale features. The deepest feature in the multi-scale features is combined with e to generate a global degradation condition vector c. A spatial reliability map R is generated based on the multi-scale features and c. Y is input into the coarse reconstruction image restoration network, and c and R are used to modulate the coarse reconstruction image restoration network to obtain the coarse reconstruction result Xc; Y, Xc and R are concatenated and input into the detail residual image restoration network, and c and R are used to modulate the detail residual image restoration network to obtain the detail residual ΔX. A gate value G is generated based on the residual energy of Y, Xc, R and ΔX. A fusion weight W is generated based on G and R. Based on Xc, the W-gated ΔX is added to Xc and restricted to a preset image value range to obtain the final enhanced image X.

2. The physical condition-guided single-pixel initial reconstruction image enhancement method according to claim 1, characterized in that, The physical condition encoder includes: The sampling rate encoding branch is used to generate a sampling rate feature map A based on the sampling rate α. α and the sampling rate embedding vector e; A physical prior extraction branch is used to compute the gradient magnitude map Gm and the Laplacian response map L in parallel from the initial reconstructed image Y; The multi-scale feature coding branch is used to perform multi-scale feature coding on the physical augmentation input P to obtain multi-scale features; The global conditional generation branch is used to concatenate the deepest feature in the multi-scale features with the sampling rate embedding vector e after global average pooling, and obtain the global degradation conditional vector c through fully connected mapping. The spatial reliability generation branch is used to upsample the features at each scale to the spatial size of the initial reconstructed image Y and then stitch them together. It is then fused with the conditional feature map obtained by expanding the global degradation conditional vector c, and the spatial reliability map R is obtained through convolution mapping and the Sigmoid function.

3. The physical condition-guided single-pixel initial reconstruction image enhancement method according to claim 2, characterized in that, The gradient magnitude map Gm is calculated using a fixed Sobel operator, and the Laplace response map L is calculated using a fixed Laplace operator; the sampling rate feature map A... α The sampling rate α is obtained by broadcasting along the spatial dimension; the sampling rate embedding vector e is obtained by sequentially connecting the first fully connected layer, the GELU activation layer, and the second fully connected layer; the multi-scale feature encoding branch includes an initial convolutional layer and a three-level stride convolutional layer to form four levels of features: original size, half size, quarter size, and eighth size.

4. The physical condition-guided single-pixel initial reconstruction image enhancement method according to claim 1, characterized in that, Both the coarse reconstruction image restoration network and the detail residual image restoration network adopt a U-shaped Transformer structure including an encoder, a bottleneck layer, and a decoder; the embedding dimension and network depth of the detail residual image restoration network are smaller than those of the coarse reconstruction image restoration network; the coarse reconstruction image restoration network is used to prioritize the restoration of the target subject contour and low-frequency structure, while the detail residual image restoration network is used to supplement edges, high-frequency textures, and local grayscale variations.

5. The physical condition-guided single-pixel initial reconstruction image enhancement method according to claim 4, characterized in that, The Transformer blocks in the coarse reconstruction image restoration network and the detail residual image restoration network employ global degradation conditional modulation and a reliability-guided window attention mechanism. The global degradation conditional modulation uses the global degradation conditional vector c to generate scaling parameter γ(c) and offset parameter β(c), and performs F′=F⊙(1+γ(c))+β(c) on the normalized feature F; F′ is the modulated feature. The reliability-guided window attention mechanism generates a position-by-position, attention-by-attention reliability score based on the input features. The reliability score is scaled and offset modulated using a global degradation condition vector c, and a reliability prior obtained by log-probability transformation of the spatial reliability map R is superimposed. The modulated reliability score is used to modulate the token features in a residual manner and to construct the pairwise attention bias in the window self-attention.

6. The physical condition-guided single-pixel initial reconstruction image enhancement method according to claim 1, characterized in that, The residual energy map E is obtained by taking the absolute value of the detail residual ΔX along the channel dimension and averaging it. Y, Xc, R and E are concatenated and input into the gating network, and the gating value G is obtained by the Sigmoid function. The fusion weight W satisfies W=clamp(λR+(1-λ)G,0,1), where λ is the fusion coefficient of the reliability map and the learned gating value. The final enhanced image satisfies X=clamp(Xc+W⊙ΔX,0,1).

7. The physical condition-guided single-pixel initial reconstruction image enhancement method according to claim 6, characterized in that, The physical condition encoder, coarse reconstruction image restoration network, detail residual image restoration network, and gating network are jointly trained end-to-end. The total loss is obtained by weighting the coarse reconstruction loss, the final reconstruction loss, the gated detail loss, the edge consistency loss, the frequency domain consistency loss, and the reliability loss. The coarse reconstruction loss and the final reconstruction loss respectively constrain the coarse reconstruction result Xc and the final enhanced image X to be close to the sharp target image T; the gated detail loss constrains W⊙ΔX to be close to the difference between the sharp target image T and the coarse reconstruction result Xc after stopping the gradient; the edge consistency loss constrains the gradient magnitude of the final enhanced image X to be consistent with that of the sharp target image T; the frequency domain consistency loss constrains the Fourier magnitude of the final enhanced image X to be consistent with that of the sharp target image T; the reliability loss constrains R to be close to the weakly supervised reliability target constructed based on the reconstruction error between the initial reconstructed image Y, the coarse reconstruction result Xc and the sharp target image T.

8. A physical condition-guided single-pixel initial reconstruction image enhancement system, characterized in that, It uses the physical condition-guided single-pixel initial reconstruction image enhancement method as described in any one of claims 1 to 7; The single-pixel initial reconstruction image enhancement system includes: The input module is used to acquire the initial reconstructed image Y of the single-pixel imaging to be enhanced and its corresponding sampling rate α; The physical conditional encoding module is used to compute the gradient magnitude map Gm and Laplace response map L of Y through the physical conditional encoder, and to expand the sampling rate α into a sampling rate feature map A with the same spatial size as Y. α And map the sampling rate α to the sampling rate embedding vector e; and map Y, Gm, L, and A α The physical augmentation input P is formed by concatenating along the channel dimension. Multi-scale feature encoding is performed on P to obtain multi-scale features. The deepest feature in the multi-scale features is combined with e to generate a global degradation condition vector c. A spatial reliability map R is generated based on the multi-scale features and c. The progressive reconstruction module is used to input Y into the coarse reconstruction image restoration network and modulate the coarse reconstruction image restoration network using c and R to obtain the coarse reconstruction result Xc; after stitching Y, Xc and R, it is input into the detail residual image restoration network and modulates the detail residual image restoration network using c and R to obtain the detail residual ΔX. The reliability fusion module is used to generate a gate value G based on the residual energy of Y, Xc, R and ΔX, and generate a fusion weight W based on G and R. Based on Xc, the W-gated ΔX is added to Xc and restricted to a preset image value range to obtain the final enhanced image X.

9. A method for identifying aerial targets, characterized in that, It includes the following steps: Acquire single-pixel imaging measurement data and corresponding sampling rates in the target spatial domain; use compressed sensing reconstruction or filtered back projection methods to form an initial reconstructed image Y; The initial reconstructed image Y is enhanced using the physical condition-guided single-pixel initial reconstructed image enhancement method as described in any one of claims 1 to 7 to obtain the final enhanced image X; The final enhanced image X is input into the target recognition model to obtain the category and / or location of the aerial target.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When executed by a processor, the computer program / instructions implement the steps of the physical condition-guided single-pixel initial reconstruction image enhancement method as described in any one of claims 1 to 7.

Images