Fourier single-pixel imaging system-based reconstruction detection frame system

Abstract

The invention discloses a reconstruction detection framework system based on a Fourier single-pixel imaging system, and the system comprises the steps: extracting features from an under-sampling image through a reconstruction network, and obtaining a Fourier single-pixel imaging reconstruction result; the reconstruction network can ensure excellent reconstruction performance and reduce the complexity of the model at the same time; a Fourier single-pixel imaging reconstruction result is sent to a target detection module for target detection, a target detection network is selected as a reference model of the detection module, a fast spatial pyramid pooling module in the target detection network is optimized into a hybrid fast spatial pyramid pooling module, and a deformable convolutional network is introduced into the target detection module; the joint training strategy module uses detection loss, guides training of a Fourier single-pixel image reconstruction network through back propagation, forces the reconstruction network to generate an image more suitable for detection, improves the final performance of target detection, and has the technical characteristics of enhancing the quality of an under-sampling image, improving the target detection performance and the like.

Description

Technical Field

[0001] The present invention relates to a reconstruction detection framework system, and more specifically, to a reconstruction detection framework system based on a Fourier single-pixel imaging system, and belongs to the field of optical imaging. Background Art

[0002] Object detection is widely used in robot navigation, intelligent video surveillance, industrial inspection, underwater detection and other fields, and is a hot topic in the field of computer vision. Since the image source of traditional object detection comes from array imaging devices, it faces many challenges under extreme environmental conditions, such as light attenuation in long-distance environments, light scattering effects, and limitations on observation angles and depths. Therefore, it is urgent to explore new imaging technologies to overcome these challenges.

[0003] Fourier single-pixel imaging is a highly robust imaging technique that uses a single-pixel detector without spatial resolution to reconstruct target scene information. Compared with traditional optical imaging, Fourier single-pixel imaging has the advantages of strong noise resistance, wide spectral band, non-local imaging, and long imaging range. Therefore, combining target detection technology with Fourier single-pixel imaging has the potential to detect targets in optically unfriendly environments, especially in large-scale dynamic scenes such as scattering media and remote sensing. However, in order to increase the speed of Fourier single-pixel imaging, the target scene needs to be undersampled, which inevitably leads to ringing effects in the imaging results. In addition, as the sampling rate decreases, the quality of the reconstruction results will also decrease. All of these disadvantages limit the subsequent target detection tasks of Fourier single-pixel imaging.

[0004] In recent years, the reconstruction algorithm of Fourier single-pixel imaging based on deep learning has alleviated the contradiction between imaging quality and sampling rate to a certain extent, making it possible to use the reconstruction results of Fourier single-pixel imaging for target detection. However, there are still some unresolved issues in the target detection task of Fourier single-pixel imaging. First, there are a large number of parameters in the current high-quality Fourier single-pixel imaging reconstruction network; after combining the large-parameter reconstruction network with the target detection network, the model is difficult to train and deploy. In addition, most reconstruction methods give priority to generating visually pleasing images, ignoring the purpose of the target detection task. The independent optimization of the Fourier single-pixel imaging reconstruction task and the target detection task leads to a gap between the reconstruction task and the detection task, which hinders the detector from achieving optimal detection performance. Therefore, it is necessary to design a joint framework that can perform image reconstruction and real-time detection. Summary of the invention

[0005] In order to solve the above-mentioned problems in the prior art, the present invention provides a reconstruction detection framework system based on a Fourier single-pixel imaging system, which can enhance the quality of undersampled images and improve target detection performance. In addition, compared with other reconstruction and detection combined methods, it is superior in detection accuracy and model efficiency, and provides new ideas and research directions for the combination of single-pixel imaging reconstruction and detection tasks.

[0006] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0007] A reconstruction detection framework system based on a Fourier single pixel imaging system, comprising an image reconstruction module, a target detection module and a joint training strategy module connected by communication;

[0008] The image reconstruction module includes a spatially adaptive reconstruction network, which extracts features from undersampled images for Fourier single-pixel imaging reconstruction to obtain Fourier single-pixel imaging reconstruction results; the image reconstruction module includes a spatial feature attention module and an adaptive feature aggregation module. The reconstruction network can significantly reduce the model complexity while ensuring excellent reconstruction performance, thereby reducing the additional computational burden imposed in the reconstruction-detection framework;

[0009] The Fourier single-pixel imaging reconstruction result is sent to the target detection module for target detection. In the detection module, the target detection network YOLOv5 is selected as the benchmark model of the detection module, and the fast spatial pyramid pooling module in the target detection network YOLOv5 is optimized into a hybrid fast spatial pyramid pooling module to enhance the extraction of global information and overcome the problem of high information loss during low sampling; and a deformable convolutional network is introduced in the target detection module to effectively alleviate the problem that the fixed convolutional layer cannot fully capture irregularly shaped objects;

[0010] The joint training strategy module uses the detection loss to guide the training of the Fourier single-pixel image reconstruction network through back-propagation, forcing the reconstruction network to generate images that are more suitable for detection, so as to further improve the final performance of target detection at low sampling rates.

[0011] Preferably, the reconstruction network is a 7-level encoder-decoder network having a structure of 4 encoder layers and 3 decoder layers, and each layer of the reconstruction network contains several spatial feature attentions so that the network can fully understand and effectively utilize spatial information for Fourier single-pixel image reconstruction.

[0012] Preferably, the reconstruction phase of the image reconstruction module is used:

[0013] Step 1) In the encoder, the image acquired by the Fourier single-pixel imaging system first undergoes a 3×3 convolution operation with a stride of 1 to adjust the channel, and then passes through a spatial feature attention module and a 3×3 convolution operation with a stride of 2 for downsampling and feature extraction. The extracted features are then used as the input of the decoder;

[0014] Step 2) In the decoder, the features are upsampled and reconstructed through pixel-shuffled 1×1 convolution operations and spatial feature attention modules. To avoid information loss, the features of each downsampling step in the encoder are connected to the corresponding decoder layer through skip connections;

[0015] Step 3) At the end of the decoder, a 3×3 convolution operation with a stride of 1 is used to resize the features into a single channel;

[0016] Finally, the pixels are added to the original image to obtain the reconstruction result of the graphics, that is, the reconstructed image.

[0017] Preferably, in step 2), since the features in the encoder and decoder are semantically different, an adaptive feature aggregation module is used at the jump connection point of the decoder to dynamically aggregate information between features at different levels and filter redundant information.

[0018] Preferably, the detection stage of the target detection module:

[0019] Step 1) Input the reconstructed image into the detection network. The target detection network YOLOv5 includes a backbone network, a neck, and several detection heads;

[0020] Step 2) In the backbone network, the global and local information of the features can be captured more comprehensively through the hybrid fast pyramid pooling module, which combines average pooling and maximum pooling to fuse global and local information during feature extraction. The hybrid fast pyramid pooling module introduces an additional average pooling branch based on the original maximum pooling branch, and similarly connects three average pooling layers;

[0021] Step 3) In the neck, a deformable convolution is introduced to improve the processing capability of deformable targets in low-sampled images; through the deformable convolution, a convolution operation whose kernel shape is dynamically adjusted to adapt to changes in the target position, size and direction in the input feature map is performed.

[0022] Preferably, the joint training strategy of the joint training strategy module is:

[0023] Step 1) In the reconstruction phase, the network used for reconstruction is trained for different sampling rates. The training process is iterated using the ADAM optimizer for 100 iterations. The batch size is set to 16, the initial learning rate is 1×10-4, and cosine annealing is used to adjust the learning rate. The best model is saved during the training process;

[0024] Step 2) In the detection phase, the FPSI reconstruction results are input into the detector for training; the training is performed using the ADAM optimizer for 200 iterations, with a batch size of 16 and a learning rate of 1×10-2, and the best model is saved during the training process;

[0025] Step 3) In the joint training phase, the weights of the reconstruction phase and the detection phase are loaded, and the reconstruction network and the detection network are trained simultaneously. The respective parameters are adjusted together through the detection loss and the reconstruction loss. Whether to end the training is decided by statistically analyzing the changes in their respective losses and the result graphs saved during each round.

[0026] Preferably, the total loss of the reconstruction detection framework system based on the Fourier single pixel imaging system is calculated as follows, where α and β are both set to 1:

[0027] L=αL rec +βL det (1)

[0028] In the reconstruction module, the loss function mainly consists of pixel loss and perceptual loss. The reconstruction module loss is shown below, and λ is set to 0.01:

[0029] L rec =L1+λL per (2)

[0030] The pixel loss L1 is expressed as:

[0031]

[0032] Where n is the number of pixels in the image, y a is the FPSI reconstruction result, y is the original image;

[0033] The perceptual loss is denoted as L per , expressed as:

[0034]

[0035] Among them, m represents the number of layers of the VGG-19 network, φ i Represents each layer used to calculate the loss in VGG-19;

[0036] In the detection module, the loss of the detection network is simply expressed as:

[0037] Ldet =α box L Clou +α obj L obj +α cls L cls (5)

[0038] where α * represents the weight of the corresponding item, L CIou represents the regression box loss, L obj represents the object confidence loss, L cls represents the category loss;

[0039] L CIou It is a regression box loss function that comprehensively considers the overlapping area, center point distance and width-to-length ratio; L CIou It can be expressed as:

[0040]

[0041]

[0042]

[0043] Among them, IOU is the intersection-over-union ratio between the detection box and the real box, ρ(b,b gt ) is the distance between the center of the detection box and the center of the real box, w and w gt are the widths of the detection box and the real box, h and h respectively. gt are the lengths of the detection box and the real box respectively. L in formula (5) obj and L cls They are all binary cross entropy functions, and the formula is as follows:

[0044]

[0045] Where y is a binary label 0 or 1 and p(y) is the probability that the output belongs to label y.

[0046] Beneficial effects: When performing target detection on low-sampling rate images, it can effectively improve the image quality and establish a closer upstream and downstream relationship to further improve the accuracy of target detection tasks; it can enhance the quality of undersampled images and improve target detection performance. In addition, compared with other reconstruction and detection combined methods, it is better in detection accuracy and model efficiency, providing new ideas and research directions for the combination of single-pixel imaging reconstruction and detection tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a reconstruction detection framework diagram of the present invention.

[0048] Figure 2It is one of the real experimental schematic diagrams of the present invention.

[0049] Figure 3 This is the second schematic diagram of the real experiment of the present invention.

[0050] Figure 4 This is the third schematic diagram of the real experiment of the present invention.

[0051] Figure 5 This is the fourth schematic diagram of the real experiment of the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings, but the present invention is not limited to the following embodiments.

[0053] The invention discloses a reconstruction detection framework system based on a Fourier single pixel imaging system, comprising an image reconstruction module and a target detection module. In the Fourier single pixel imaging reconstruction module, a high-performance spatial adaptive reconstruction network is designed. The network is intended to extract features from undersampled images for Fourier single pixel imaging reconstruction. The network is mainly composed of a spatial feature attention module and an adaptive feature aggregation module. While ensuring excellent reconstruction performance, the model complexity is significantly reduced, thereby reducing the additional computational burden imposed in the reconstruction-detection framework. Then, the Fourier single pixel imaging reconstruction result is sent to the target detection module for target detection. In the target detection module, the YOLOv5 network is improved. Among them, a hybrid fast spatial pyramid pooling structure is proposed, which enhances the extraction of global information compared with the original fast spatial pyramid pooling structure. In addition, a deformable convolutional network is introduced to effectively alleviate the problem that a fixed convolutional layer cannot fully capture irregularly shaped objects. Finally, in order to better assist subsequent target detection tasks, a joint training strategy module for the reconstruction network and the target detection network is designed. The strategy aims to further improve the target detection performance at low sampling rates. Numerical simulation and real experimental results show that the proposed Fourier single-pixel imaging reconstruction-detection framework can enhance the quality of undersampled images and improve target detection performance. In addition, compared with other reconstruction and detection methods, it is superior in detection accuracy and model efficiency. It provides new ideas and research directions for the combination of single-pixel imaging reconstruction and detection tasks:

[0054] 1) In the FSPI reconstruction stage, a spatial information attention module and an adaptive feature aggregation module are designed, which significantly enhance the network's ability to extract multi-scale features and suppress noise, and effectively reduce the number of parameters of the reconstruction network.

[0055] 2) In the target detection stage, the fast spatial pyramid pooling module is improved to a hybrid fast spatial pyramid pooling module to enhance its ability to extract global information. In addition, in order to alleviate the impact of the ringing effect on target deformation, a deformable convolution is introduced.

[0056] 3) We design a joint training strategy that uses the detection loss to guide the training of the Fourier single-pixel image reconstruction network through back-propagation. This forces the reconstruction network to generate images that are more suitable for detection, further improving the final performance of object detection at low sampling rates.

[0057] like Figure 1-5 The figure shows a specific embodiment of a reconstruction detection framework system based on a Fourier single pixel imaging system, which embodiment is a reconstruction detection framework system based on a Fourier single pixel imaging system, including an image reconstruction module, a target detection module and a joint training strategy module connected in communication;

[0058] The image reconstruction module includes a spatially adaptive reconstruction network, which extracts features from undersampled images for Fourier single-pixel imaging reconstruction to obtain Fourier single-pixel imaging reconstruction results; the image reconstruction module includes a spatial feature attention module and an adaptive feature aggregation module. The reconstruction network can significantly reduce the model complexity while ensuring excellent reconstruction performance, thereby reducing the additional computational burden imposed in the reconstruction-detection framework;

[0059] The Fourier single-pixel imaging reconstruction result is sent to the target detection module for target detection. In the detection module, the target detection network YOLOv5 is selected as the benchmark model of the detection module, and the fast spatial pyramid pooling module in the target detection network YOLOv5 is optimized into a hybrid fast spatial pyramid pooling module to enhance the extraction of global information and overcome the problem of high information loss during low sampling; and a deformable convolutional network is introduced in the target detection module to effectively alleviate the problem that the fixed convolutional layer cannot fully capture irregularly shaped objects;

[0060] The joint training strategy module uses the detection loss to guide the training of the Fourier single-pixel image reconstruction network through back-propagation, forcing the reconstruction network to generate images that are more suitable for detection, so as to further improve the final performance of target detection at low sampling rates.

[0061] 1. The reconstruction network (reconstruction network) structure is a 7-level encoder-decoder network with 4 layers of encoder and 3 layers of decoder. Each layer of the spatial adaptive reconstruction network contains several spatial feature attentions, which enables the network to fully understand and effectively use spatial information for Fourier single-pixel image reconstruction. In the encoder, the image first undergoes a 3×3 convolution operation with a stride of 1 to adjust the channel. Then, it passes through the spatial information attention module and a 3×3 convolution operation with a stride of 2 for downsampling and feature extraction. The extracted features are then used as input to the decoder. In the decoder, these features are upsampled and reconstructed through pixel shuffling and 1×1 convolution operations and spatial information attention modules. To avoid information loss, the features of each downsampling step in the encoder are connected to the corresponding decoder layer through skip connections. Since the features in the encoder and decoder are semantically different, an adaptive feature aggregation module is used at the skip connection points of the decoder. The adaptive feature aggregation module dynamically aggregates information between features at different levels and filters redundant information. At the end of the decoder, a 3×3 convolution operation with a stride of 1 is used to adjust the features to a single channel. Finally, they are added pixel by pixel to the original image to obtain the final reconstruction result.

[0062] 2. YOLOv5 is selected as the benchmark model of the detection module. In order to solve the problem that the fast spatial pyramid pooling module in YOLOv5 only uses maximum pooling, resulting in a lot of information loss when undersampling, a hybrid fast spatial pyramid pooling module is designed. In addition, in order to solve the problem of object distortion caused by undersampling, deformable convolution is introduced into the neck.

[0063] 3. There is a fundamental difference between Fourier single-pixel image reconstruction and target detection tasks because their goals are not exactly the same. The main goal of Fourier single-pixel image reconstruction is to enhance image quality, which is usually measured by indicators such as peak signal-to-noise and structural similarity index. In contrast, the purpose of target detection is to improve the accuracy of machine vision, which is usually evaluated using indicators such as the average accuracy of all categories. In order to solve these different goals, the trained tasks are combined, and the back-propagated detection loss is used to guide the training of the Fourier single-pixel image reconstruction network, forcing it to generate images that are more suitable for detection. The specific operations of the reconstruction detection method based on Fourier single-pixel imaging are as follows: Figure 2 As shown:

[0064] Step 1: In the reconstruction phase, the network used for reconstruction is trained for different sampling rates. The training process is iterative using the ADAM optimizer and requires about 100 iterations. The batch size is set to 16, the initial learning rate is 1×10-4, and cosine annealing is used to adjust the learning rate.

[0065] Step 2: In the detection phase, the FPSI reconstruction results are input into the detector for training. Similarly, the ADAM optimizer is used for training, which takes about 200 iterations, with a batch size of 16 and a learning rate of 1×10-2. The best model is saved during the training process.

[0066] Step 3: In the joint training phase, load the pre-trained weights and train the reconstruction network and the detection network simultaneously to fine-tune their respective parameters. Decide whether to end the training by observing the changes in their respective losses and the result graphs saved during each round.

[0067] In a preferred embodiment, the reconstruction network is a 7-level encoder-decoder network with a structure of 4 encoder layers and 3 decoder layers. Each layer of the reconstruction network contains several spatial feature attentions so that the network can fully understand and effectively utilize spatial information for Fourier single-pixel image reconstruction.

[0068] A preferred embodiment uses the reconstruction phase of the image reconstruction module:

[0069] Step 1) In the encoder, the image acquired by the Fourier single-pixel imaging system first undergoes a 3×3 convolution operation with a stride of 1 to adjust the channel, and then passes through a spatial feature attention module and a 3×3 convolution operation with a stride of 2 for downsampling and feature extraction. The extracted features are then used as the input of the decoder;

[0070] Step 2) In the decoder, the features are upsampled and reconstructed through pixel-shuffled 1×1 convolution operations and spatial feature attention modules. To avoid information loss, the features of each downsampling step in the encoder are connected to the corresponding decoder layer through skip connections;

[0071] Step 3) At the end of the decoder, a 3×3 convolution operation with a stride of 1 is used to resize the features into a single channel;

[0072] Finally, the pixels are added to the original image to obtain the reconstruction result of the graphics, that is, the reconstructed image.

[0073] In a preferred embodiment, in step 2), since the features in the encoder and the decoder are semantically different, an adaptive feature aggregation module is used at the jump connection point of the decoder to dynamically aggregate information between features at different levels and filter out redundant information.

[0074] In a preferred embodiment, the detection phase of the target detection module is:

[0075] Step 1) Input the reconstructed image into the detection network. The target detection network YOLOv5 includes a backbone network, a neck, and several detection heads;

[0076] Step 2) In the backbone network, the global and local information of the features can be captured more comprehensively through the hybrid fast pyramid pooling module, which combines average pooling and maximum pooling to fuse global and local information during feature extraction. The hybrid fast pyramid pooling module introduces an additional average pooling branch based on the original maximum pooling branch, and similarly connects three average pooling layers;

[0077] Step 3) In the neck, a deformable convolution is introduced to improve the processing capability of deformable targets in low-sampled images; through the deformable convolution, a convolution operation whose kernel shape is dynamically adjusted to adapt to changes in the target position, size and direction in the input feature map is performed.

[0078] In a preferred embodiment, the joint training strategy of the joint training strategy module is:

[0079] Step 1) In the reconstruction phase, the network used for reconstruction is trained for different sampling rates. The training process is iterated using the ADAM optimizer for 100 iterations. The batch size is set to 16, the initial learning rate is 1×10-4, and cosine annealing is used to adjust the learning rate. The best model is saved during the training process;

[0080] Step 2) In the detection phase, the FPSI reconstruction results are input into the detector for training; the training is performed using the ADAM optimizer for 200 iterations, with a batch size of 16 and a learning rate of 1×10-2, and the best model is saved during the training process;

[0081] Step 3) In the joint training phase, the weights of the reconstruction phase and the detection phase are loaded, and the reconstruction network and the detection network are trained simultaneously. The respective parameters are adjusted together through the detection loss and the reconstruction loss. Whether to end the training is decided by statistically analyzing the changes in their respective losses and the result graphs saved during each round.

[0082] Preferably, the total loss of the reconstruction detection framework system based on the Fourier single pixel imaging system is calculated as follows, where α and β are both set to 1:

[0083] L=αL rec +βL det (10)

[0084] In the reconstruction module, the loss function mainly consists of pixel loss and perceptual loss. The reconstruction module loss is shown below, and λ is set to 0.01:

[0085] L rec =L1+λL per (11)

[0086] The pixel loss L1 is expressed as:

[0087]

[0088] Where n is the number of pixels in the image, y a is the FPSI reconstruction result, y is the original image;

[0089] The perceptual loss is denoted as L per , expressed as:

[0090]

[0091] Among them, m represents the number of layers of the VGG-19 network, φ i Represents each layer used to calculate the loss in VGG-19;

[0092] In the detection module, the loss of the detection network is simply expressed as:

[0093] L det =α box L Clou +α obj L obj +α cls L cls (14)

[0094] where α * represents the weight of the corresponding item, L CIou represents the regression box loss, L obj represents the object confidence loss, L cls represents the category loss;

[0095] L CIou It is a regression box loss function that comprehensively considers the overlapping area, center point distance and width-to-length ratio; L CIou It can be expressed as:

[0096]

[0097]

[0098]

[0099] Among them, IOU is the intersection-over-union ratio between the detection box and the real box, ρ(b,b gt ) is the distance between the center of the detection box and the center of the real box, w and w gt are the widths of the detection box and the real box, h and h respectively. gt are the lengths of the detection box and the real box respectively. L in formula (5) obj and L cls They are all binary cross entropy functions, and the formula is as follows:

[0100]

[0101] Where y is a binary label 0 or 1 and p(y) is the probability that the output belongs to label y.

[0102] The present invention uses joint training-based image reconstruction and target detection to improve the accuracy of current low-sampling rate Fourier single-pixel image target detection. In order to handle the reconstruction task of low-sampling rate Fourier single-pixel images with low quality, a spatial adaptive reconstruction network is designed to quickly reconstruct low-sampling rate images to improve image quality. Hybrid fast spatial pooling pyramid fusion and deformable volume are fused into the target detection network to improve detection performance. A joint training strategy is designed to strengthen the synergy between high-level and low-level visual tasks, thereby further improving detection accuracy.

[0103] Finally, it should be noted that the present invention is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or associated with the content disclosed by ordinary technicians in this field should be considered as the protection scope of the present invention.

Claims

1. A reconstruction detection framework system based on Fourier single pixel imaging system, characterized by: It includes a communication-connected image reconstruction module, a target detection module, and a joint training strategy module; The image reconstruction module includes a spatially adaptive reconstruction network, which extracts features from undersampled images for Fourier single-pixel imaging reconstruction to obtain Fourier single-pixel imaging reconstruction results; the image reconstruction module includes a spatial feature attention module and an adaptive feature aggregation module. The reconstruction network can significantly reduce the model complexity while ensuring excellent reconstruction performance, thereby reducing the additional computational burden imposed in the reconstruction-detection framework; The Fourier single-pixel imaging reconstruction result is sent to the target detection module for target detection. In the detection module, the target detection network YOLOv5 is selected as the benchmark model of the detection module, and the fast spatial pyramid pooling module in the target detection network YOLOv5 is optimized into a hybrid fast spatial pyramid pooling module to enhance the extraction of global information and overcome the problem of high information loss during low sampling. In the object detection module, a deformable convolutional network is introduced to effectively alleviate the problem that fixed convolutional layers cannot fully capture irregularly shaped objects. The joint training strategy module uses the detection loss to guide the training of the Fourier single-pixel image reconstruction network through back-propagation, forcing the reconstruction network to generate images that are more suitable for detection, so as to further improve the final performance of target detection at low sampling rates.

2. The reconstruction detection framework system based on Fourier single pixel imaging system according to claim 1, characterized in that: The reconstruction network is a 7-level encoder-decoder network with a 4-layer encoder and 3-layer decoder structure. Each layer of the reconstruction network contains several spatial feature attentions so that the network can fully understand and effectively utilize spatial information for Fourier single-pixel image reconstruction.

3. The reconstruction detection framework system based on Fourier single pixel imaging system according to claim 1, characterized in that: Reconstruction phase using the image reconstruction module: Step 1) In the encoder, the image acquired by the Fourier single-pixel imaging system first undergoes a 3×3 convolution operation with a stride of 1 to adjust the channel, and then passes through a spatial feature attention module and a 3×3 convolution operation with a stride of 2 for downsampling and feature extraction. The extracted features are then used as the input of the decoder; Step 2) In the decoder, the features are upsampled and reconstructed through pixel-shuffled 1×1 convolution operations and spatial feature attention modules. To avoid information loss, the features of each downsampling step in the encoder are connected to the corresponding decoder layer through skip connections; Step 3) At the end of the decoder, a 3×3 convolution operation with a stride of 1 is used to resize the features into a single channel; Finally, the pixels are added to the original image to obtain the reconstruction result of the graphics, that is, the reconstructed image.

4. The reconstruction detection framework system based on Fourier single pixel imaging system according to claim 3, characterized in that: In step 2), since the features in the encoder and decoder are semantically different, an adaptive feature aggregation module is used at the jump connection point of the decoder to dynamically aggregate information between features at different levels and filter out redundant information.

5. The reconstruction detection framework system based on Fourier single pixel imaging system according to claim 1, characterized in that: Detection phase of the target detection module: Step 1) Input the reconstructed image into the detection network. The target detection network YOLOv5 includes a backbone network, a neck, and several detection heads; Step 2) In the backbone network, the global and local information of the features can be captured more comprehensively through the hybrid fast pyramid pooling module, which combines average pooling and maximum pooling to fuse global and local information during feature extraction. The hybrid fast pyramid pooling module introduces an additional average pooling branch based on the original maximum pooling branch, and similarly connects three average pooling layers; Step 3) In the neck, a deformable convolution is introduced to improve the processing capability of deformable targets in low-sampled images; through the deformable convolution, a convolution operation whose kernel shape is dynamically adjusted to adapt to changes in the target position, size and direction in the input feature map is performed.

6. The reconstruction detection framework system based on Fourier single pixel imaging system according to claim 4, characterized in that: Joint training strategy module's joint training strategy: Step 1) In the reconstruction phase, the network used for reconstruction is trained for different sampling rates. The training process is iterated using the ADAM optimizer for 100 iterations. The batch size is set to 16, the initial learning rate is 1×10-4, and cosine annealing is used to adjust the learning rate. The best model is saved during the training process; Step 2) In the detection phase, the FPSI reconstruction results are input into the detector for training; The training was performed using the ADAM optimizer for 200 iterations, a batch size of 16, a learning rate of 1×10-2, and the best model was saved during the training process; Step 3) In the joint training phase, the weights of the reconstruction phase and the detection phase are loaded, and the reconstruction network and the detection network are trained simultaneously. The respective parameters are adjusted together through the detection loss and the reconstruction loss. Whether to end the training is decided by statistically analyzing the changes in their respective losses and the result graphs saved during each round.

7. A reconstruction detection framework system based on Fourier single pixel imaging system according to any one of claims 1 to 6, characterized in that: The total loss of the reconstruction detection framework system based on the Fourier single-pixel imaging system is calculated as follows, where α and β are both set to 1: L=αL rec +βL det (1) In the reconstruction module, the loss function mainly consists of pixel loss and perceptual loss. The reconstruction module loss is shown below, and λ is set to 0.01: L rec =L1+λL per (2) The pixel loss L1 is expressed as: Where n is the number of pixels in the image, y a is the FPSI reconstruction result, y is the original image; The perceptual loss is denoted as L per , expressed as: Among them, m represents the number of layers of the VGG-19 network, φ i Represents each layer used to calculate the loss in VGG-19; In the detection module, the loss of the detection network is simply expressed as: L det =a box L Clou +a obj L obj +a cls L cls (5) where α * represents the weight of the corresponding item, L CIou represents the regression box loss, L obj represents the object confidence loss, L cls represents the class loss; L CIou It is a regression box loss function that comprehensively considers the overlapping area, center point distance and width-to-length ratio; L CIou It can be expressed as: Among them, IOU is the intersection-over-union ratio between the detection box and the real box, ρ(b,b gt ) is the distance between the center of the detection box and the center of the real box, w and w gt are the widths of the detection box and the real box, h and h respectively. gt are the lengths of the detection box and the real box respectively. L in formula (5) obj and L cls They are all binary cross entropy functions, and the formula is as follows: Where y is a binary label 0 or 1 and p(y) is the probability that the output belongs to label y.

Images