A zero-shot pedestrian re-identification method and system in low-light environments

Abstract

This invention discloses a zero-shot pedestrian re-identification method and system in low-light environments, relating to the field of computer vision technology. The method includes the following steps: generating a low-light image with missing illumination to obtain image pair 1; extracting features from image pair 1 using a feature encoder; processing the features using a transposed convolutional decoder to obtain a decoding result, and then processing the features using a re-identification network to obtain a re-identification result; decomposing the images in image pair 1 into reflection and illumination components; swapping the reflection components of image pair 1, and then recombining the swapped reflection components with the original illumination components to obtain image pair 2; processing image pair 2 using a feature encoder and a transposed convolutional decoder to obtain a decoding result; calculating a loss function based on the decoding result of image pair 1, the re-identification result, and the decoding result of image pair 2; and optimizing the network parameters of the re-identification network based on the loss function. This invention improves the accuracy of re-identification.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, and in particular relates to a method and system for pedestrian re-identification in low-light environments with zero samples. Background Technology

[0002] With rapid urban development and increasing public awareness of safety, pedestrian re-identification technology is becoming increasingly important in fields such as video surveillance, intelligent traffic management, and security systems. Traditional pedestrian re-identification technologies typically rely on well-lit environments, where cameras can capture clear images, leading to more accurate pedestrian identification. However, in low-light environments, such as at night, dimly lit indoor spaces, and underground parking lots, pedestrian images are often blurry, making existing pedestrian re-identification methods inadequate.

[0003] To address the challenges of pedestrian recognition in low-light environments, pedestrian re-identification techniques for low-light environments have been developed in recent years. These techniques typically rely on training datasets containing low-light images and employ specific image enhancement and processing methods to improve recognition accuracy. However, in practical applications, it is often impossible to obtain enough low-light images for training. This leads to a greater challenge for pedestrian re-identification techniques when the training set only contains well-lit images and lacks zero-sample low-light images.

[0004] Zero-shot pedestrian re-identification technology in low-light environments aims to solve this challenge: effectively identifying pedestrians in low-light conditions without the availability of low-light images as training samples. The core of this technology lies in how to utilize images acquired only under good lighting conditions, and through algorithm design and network model optimization, enable the recognition system to adapt to low-light conditions and accurately identify pedestrians. This method has broad application prospects, especially in scenarios where obtaining low-light images is difficult, such as all-weather surveillance systems and nighttime pedestrian detection in intelligent traffic management.

[0005] Currently, existing pedestrian re-identification methods in low-light environments typically rely on multimodal data, such as combining visible light and infrared images, to improve recognition accuracy under low-light conditions. However, these methods are not only complex to acquire data for, but also difficult to implement in practical applications. Pedestrian re-identification technology for zero-shot scenarios in low-light environments is still in the early research stages, and there are no mature commercial solutions yet. Therefore, developing a pedestrian re-identification technology based solely on well-lit images, capable of maintaining high accuracy under low-light conditions, has become a major challenge and hot topic in current technological research and development.

[0006] Similarly, no patents related to zero-shot pedestrian re-identification technology in low-light environments were found. Some patents, such as the frame extraction pedestrian retrieval method based on YOLOv5s and staged regularization joint pedestrian re-identification (publication number CN118522039A), have achieved significant progress by combining pedestrian re-identification algorithms and target detection technology. However, in low-light applications, several problems remain to be solved, such as image quality degradation under low light conditions, reduced accuracy of pedestrian feature extraction, and insufficient model generalization ability. The pedestrian re-identification method and device based on attention mechanisms and feature fusion strategies (publication number CN118470752A) improves the accuracy of pedestrian re-identification by introducing attention mechanisms and multimodal feature fusion strategies. However, these methods still rely on large-scale labeled datasets and struggle to handle zero-shot pedestrian identification. Existing methods do not address pedestrian re-identification in zero-shot and low-light environments.

[0007] This invention proposes a zero-sample pedestrian re-identification method and system in low-light environments, thereby solving the current problem of pedestrian re-identification in low-light environments. Summary of the Invention

[0008] The purpose of this invention is to provide a pedestrian re-identification method and system in low-light environments with zero samples, so as to solve the problem mentioned in the background art that the existing technology for pedestrian re-identification in low-light environments lacks research on zero-sample cases.

[0009] To achieve the above objectives, the present invention employs the following technical solution:

[0010] The first aspect of this invention provides a zero-sample pedestrian re-identification method in low-light environments, comprising the following steps:

[0011] S1. Using a dark light image synthesis network, well-lit images are processed to generate dark light images with missing lighting. The well-lit images and dark light images with missing lighting are combined to obtain image pair 1.

[0012] S2. Extract features from image pair 1 using a feature encoder;

[0013] S3. The features are processed by the transposed convolutional decoder of the reflection layer to obtain the decoding result, and the features are processed by the re-identification network to obtain the re-identification result.

[0014] S4. Use the Retinex decomposition network to decompose the image in image pair 1 into reflection component and illumination component respectively.

[0015] S5. Swap the reflection components of image pair 1, and then resynthesize the swapped reflection components with the original illumination components to obtain image pair 2.

[0016] S6. The image pair 2 is processed by the feature encoder to obtain the features, and the features are decoded by the transposed convolutional decoder of the reflection layer to obtain the decoding result.

[0017] S7. Calculate the loss function based on the decoding and re-identification results of image pair 1 and image pair 2. The loss function includes the decoding loss. Reconstruction loss Reflector loss Triple loss and cross-entropy loss

[0018] S8. Optimize the re-identification network by adjusting the network parameters according to the loss function.

[0019] Preferably, in S1, image pair 1 (I1, I2) is obtained, specifically as follows:

[0020] I2 = DarkISP(I1)

[0021] Where I1 is a well-lit image, I2 is a dark image with missing lighting, and DarkISP is a dark image synthesis network; images that are essentially the same but have different lighting information are set as a pair, called image pair 1 (I1, I2).

[0022] Furthermore, the dark light image synthesis network generates dark light images with missing illumination, specifically as follows:

[0023] Read the well-lit image (img) and perform inverse tone mapping on it, as shown in the following formula;

[0024] img1=0.5-sin(arcsin(1.0-2.0×img) / 3.0)

[0025] The image is subjected to inverse Gamma correction, as shown in the following formula;

[0026] img2 = max(img1, ∈) γ

[0027] The Gamma value γ is randomly selected within a preset range;

[0028] Convert the image from the standard sRGB color space to the camera's RGB color space as shown in the following formula;

[0029] img3=apply_ccm(img2,rgb2cam)

[0030] The apply_ccm function is used to linearly transform the color channel of each pixel in the input image with a specified color correction matrix, thereby achieving color space conversion.

[0031] Perform reverse white balance adjustment on the image, as shown in the following formula;

[0032] img4 = img3 × gains1

[0033] Where gains1 is the gain factor for the three channels of the image;

[0034] Adjust the overall brightness of the image as shown in the following formula;

[0035] img5 = img4 × darkness

[0036] Where darkness is the darkness coefficient;

[0037] Add noise to the image as shown in the following formula;

[0038] img6 = img5 + noise1 + noise2

[0039] Where noise1 is the shooting noise and noise2 is the quantization noise;

[0040] Adjust the white balance of the image again, as shown in the following formula;

[0041] img7 = img6 × gains2

[0042] Where gains2 is the gain factor for the three channels of the image;

[0043] Convert the image from the camera's RGB color space back to the standard sRGB color space, as shown in the following formula;

[0044] img8=apply_ccm(img7,rgb2cam)

[0045] The image is subjected to Gamma correction as shown in the following formula;

[0046]

[0047] Through the above steps, the dark light image synthesis network synthesizes the well-lit image img, i.e., I1, into a dark light image img9, i.e., I2, which has low lighting characteristics and lacks illumination.

[0048] Preferably, step S2 is as follows:

[0049] F i =FeatureEncoder(I i )

[0050] Wherein, Feature Encoder is the feature encoder, I i For image pair 1(I1, I2), Fi The features output by the feature encoder.

[0051] Preferably, step S3 is as follows:

[0052]

[0053] F = ReIDNet(F) i )

[0054] Among them, TransConvRefDecoder is a transposed convolutional decoder of the reflection layer, and ReIDNet is a re-identification network; The decoding result of image pair 1 F represents the re-identification result.

[0055] Furthermore, the transposed convolutional decoder of the reflection layer processes the data to obtain the decoding result, as follows:

[0056] The encoder's encoded result is input into Conv3 and activated by the ReLU function, as shown in the following equation;

[0057] DF1 = ReLU(Conv3(Feat))

[0058] Where Feat is the output of the feature encoder;

[0059] The output DF1 of the convolutional layer Conv3 is input into the transposed convolutional layer TransConv and activated by the Sigmoid function, as shown in the following equation;

[0060]

[0061] in, The decoding result obtained after decoding, i.e., the reflection component.

[0062] Preferably, in step S4, the Retinex decomposition network is used to decompose image I1 into R1 and L1, and similarly, image I2 is decomposed into R2 and L2, as follows:

[0063] R i ,L i =RetinexDecomNet(I i )

[0064] Among them, Retinex DecomNet is the Retinex decomposition network, R i For the reflection component carrying reflection information, L i The illumination component contains illuminance information;

[0065] Furthermore, the Retinex decomposition network performs image decomposition as follows:

[0066] The maximum value channel expansion process is performed on the input image Img to extract the maximum value channel input_max of the input image, and then it is concatenated with the original image to generate an image input_img containing more information; as shown in the following formula;

[0067] input_max=torch.max(Img,dim=1)[0]

[0068] input_img=torch.cat((input_max,Img),dim=1)

[0069] Shallow feature extraction is performed on the expanded image, and preliminary features feats0 are obtained through the convolutional layer net1_conv0, as shown in the following equation;

[0070] feats0=net1_conv0(input_img)

[0071] The initial features are deeply extracted by combining multiple convolutional layers and activation functions net1_convs to generate high-level feature features, as shown in the following formula;

[0072] featss = net1_convs(feats0)

[0073] The structure of net1_convs consists of five stacks of convolutional layers and ReLU activation functions;

[0074] The high-level features are reconstructed by applying a convolutional layer net1_recon to convert the high-level features into the output channel image outs; as shown in the following formula;

[0075] outs = net1_recon(featss)

[0076] The output channel image is normalized using a sigmoid function to generate the reflection component R and the illumination component L, as shown in the following formula;

[0077] R=sigmoid(outs[:,0:3,∶,∶])

[0078] L=sigmoid(outs[:,3:4,∶,∶])

[0079] Through the above steps, the input image is decomposed into a reflection image R and an illumination image L.

[0080] Preferably, in step S5, image pair 2 (I′1, I′2) is obtained as follows:

[0081] (R1,L1),(R2,L2)→(R2,L1),(R1,L2)

[0082] I′1=R2·L1

[0083] I′2=R1·L2.

[0084] Preferably, the decoding results R′1 and R′2 obtained in step S6 are as follows:

[0085] F′ i =FeatureEncoder(I′) i )

[0086] R′ i =TransConvRefDecoder(F′ i )

[0087] Among them, I′ i For image pair 2(I′1, I′2), F′ i R′ is the feature output by the feature encoder. i Let R′1 and R′2 be the decoding results obtained from image pair 2.

[0088] Preferably, the loss function in S7 is as follows:

[0089]

[0090] Among them, In this context, MSE represents mean squared error, and SSIM represents the structural similarity index measure.

[0091] exist In this context, MAE represents the mean absolute error, and (R,R′) represents (R1,R′2) or (R2,R′1), respectively.

[0092] A second aspect of the present invention provides a zero-sample pedestrian re-identification system in low-light environments, comprising:

[0093] A low-light image synthesis network is used to generate low-light images with missing lighting.

[0094] The feature encoder consists of two convolutional layers used to encode image features;

[0095] A transposed convolutional decoder with a reflective layer is used to decode encoded features;

[0096] The Retinex decomposition network is used to decompose image pairs into illumination and reflection components.

[0097] The pedestrian re-identification network uses a VGG16 network as its backbone to obtain pedestrian re-identification results in images.

[0098] Compared with the prior art, the beneficial effects of the present invention are:

[0099] (1) This invention automatically generates low-light images to form image pairs with good lighting and poor lighting. These image pairs are then processed through two branch networks: first, a feature encoder extracts features, which are then simultaneously input into a transposed convolutional decoder and a re-identification network. The transposed convolutional decoder obtains the decoding result, and the re-identification network obtains the re-identification result. Second, a Retinex decomposition network extracts the illumination and reflection components, swaps the reflection components of the image pairs, and re-synthesizes new image pairs. These new image pairs are then input again into the feature encoder and transposed convolutional decoder to obtain new decoding results. Various loss functions are calculated based on the decoding results. These loss functions enable the network to better learn features and improve the accuracy of re-identification. Addressing the limited design of pedestrian re-identification methods for low-light images in existing technologies, this invention proposes a more novel solution specifically suitable for pedestrian re-identification in low-light environments.

[0100] (2) The network training of the present invention does not rely on pedestrian images in low light environment, which enables the network to perform re-recognition reasoning directly on low light images, improving adaptability and practicality; the present invention can effectively simulate the image characteristics under low light conditions through the low light image synthesis network, and generate realistic low light images under different lighting conditions.

[0101] (2) Based on Retinex theory, this invention designs a novel transposed convolutional decoding module for the reflection layer, which greatly improves the stability and accuracy of the network when decoding reflection information. Attached Figure Description

[0102] Figure 1 This is a flowchart of the zero-sample pedestrian re-identification method in low-light environments in this invention;

[0103] Figure 2 This is a block diagram of the overall structure of the pedestrian re-identification system in low-light environment with zero samples in this invention. Detailed Implementation

[0104] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only 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.

[0105] Example 1:

[0106] See Figure 1 The method for pedestrian re-identification in low-light environments with zero samples is implemented as follows:

[0107] Step S110: The input well-lit image is processed using the DarkISP network to generate a pair of well-lit and poorly lit images 1 (I1, I2).

[0108] like Figure 2 As shown in the DarkISP module, this embodiment specifically includes:

[0109] Step S110-1: Read the well-lit image (img);

[0110] Step S110-2: Perform inverse tone mapping on the image to restore the original tone-mapped image to an approximate RAW format. As shown in formula (1);

[0111] img1=0.5-sin(arcsin(1.0-2.0×img) / 3.0)(1)

[0112] Step S110-3: Perform inverse Gamma correction on the image to restore it to linear illumination space. The Gamma value γ is randomly selected within a preset range, as shown in formula (2);

[0113] img2 = max(img1, ∈) γ (2)

[0114] Step S110-4: After completing the inverse Gamma correction, apply the Color Correction Matrix (CCM) to convert the image from the standard sRGB color space to the camera's RGB color space (cRGB), as shown in Equation (3);

[0115] img3=apply_ccm(img2,rgb2cam)(3)

[0116] The `apply_ccm` function performs a linear transformation on the color channel of each pixel in the input image using a specified color correction matrix, thereby achieving color space conversion. The `rgb2cam` matrix converts standard display colors to the colors actually captured by the camera sensor, simulating the real shooting process.

[0117] Step S110-5: After completing the inverse processing, perform inverse white balance adjustment on the image. Adjust the gain coefficients (gains1) of the red and blue channels of the image by applying the corresponding gain values, as shown in formula (4).

[0118] img4 = img3 × gains1(4)

[0119] Here, gains1 is the gain matrix (or vector) used to perform the white balance operation. It is a three-channel gain factor used to adjust the brightness of the red, green, and blue channels in the image to simulate the effect of white balance.

[0120] Step S110-6: Adjust the overall brightness of the image to simulate brightness decay in low-light environments. This is done by randomly selecting a darkness coefficient, as shown in formula (5).

[0121] img5 = img4 × darkness(5)

[0122] Step S110-7: In order to get closer to the real low light image quality, shooting noise (noise1) is added to the image and quantization noise (noise2) is added to simulate the subtle distortion in the image digitization process, as shown in formula (6).

[0123] img6 = img5 + noise1 + noise2(6)

[0124] Step S110-8: In order to simulate the color performance of the actual captured image, the white balance of the image is adjusted again by applying the corresponding gain value to adjust the gain coefficient (gains2) of the red and blue channels, as shown in formula (7).

[0125] img7 = img6 × gains2(7)

[0126] Step S110-9: After restoring the white balance, apply the inverse CCM operation to convert the image from the camera's RGB color space (cRGB) back to the standard sRGB color space, as shown in formula (8).

[0127] img8=apply_ccm(img7,rgb2cam)(8)

[0128] Step S110-10, finally, perform Gamma correction to adjust the image to a visualized linear illumination space. The final Gamma correction is shown in formula (9).

[0129]

[0130] Through the above steps, the well-lit input image img, i.e., I1, is synthesized into an image img9, i.e., I2, with low-light characteristics. This low-light image degradation network can effectively simulate the image characteristics under dark conditions, generating realistic low-light images under different lighting conditions, providing richer and more realistic input data for model training.

[0131] Step S120: Extract features from image pair 1 using a feature encoder to obtain features.

[0132] like Figure 2 As shown in the Feature Encoder module, in this embodiment:

[0133] Step S120-1: Input image I into convolutional layer Conv1 as shown in formula (10), where the input has 3 channels, the output has 64 channels, and the kernel size is 3.

[0134] F1=Conv1(I)(10)

[0135] Step S120-2: Input the output of convolutional layer Conv1 into convolutional layer Conv2, as shown in formula (11), where the input has 64 channels, the output has 64 channels, the kernel size is 3, and Feat is the encoder input result;

[0136] Feat=Conv2(F1)(11)

[0137] Image decomposition is considered a low-level task, as it can efficiently encode low-level information of an image using only two convolutional layers, thus laying a solid foundation for subsequent reflection information decoding.

[0138] Step S130: Input the features into the transposed convolutional decoder (TransConvRefDecoder) of the reflection layer to obtain the decoding result. At the same time, the features are input into the re-identification network to obtain the re-identification result;

[0139] like Figure 2 The module shown in TransConvRefDecoder, in this embodiment, specifically includes:

[0140] Step S130-1: Input the encoder encoding result into Conv3 and activate it through the ReLU function, as shown in formula (12), where the input has 64 channels, the output has 64 channels, and the kernel size is 3.

[0141] DF1 = ReLU(Conv3(Feat))(12)

[0142] Step S130-2: Input the output of the convolutional layer Conv3 into the transposed convolutional layer TransConv, and activate it with the Sigmoid function as shown in formula (13), where the input has 64 channels, the output has 3 channels, and the kernel size is 3.

[0143]

[0144] The reflection components obtained after decoding are respectively

[0145] Step S130-3: Input the encoder encoding result into the re-identification network with VGG16 network as the backbone, as shown in formula (14), where F is the output result of the network.

[0146] F = VGG16(Feat)(14)

[0147] A transposed convolutional layer is a layer that learns the optimal upsampling method by optimizing weights through backpropagation to adapt to specific data features. This means the network can more effectively learn how to reconstruct high-resolution reflection components, potentially improving the detail and accuracy of the final image and significantly enhancing image quality.

[0148] Step S140: Use the Retinex DecomNet network to decompose the image pair into reflection components R1, R2 and illumination components L1, L2.

[0149] like Figure 2 The module shown in Retinex DecomNet, specifically in this embodiment:

[0150] Step S140-1: Perform maximum value channel expansion processing on the input image Img, extract the maximum value channel input_max of the input image, and then concatenate it with the original image to generate an image input_img containing more information, as shown in formulas (15) and (16);

[0151] input_max=torch.max(Img,dim=1)[0](15)

[0152] input_img=torch.cat((input_max,Img),dim=1)(16)

[0153] Step S140-2: Perform shallow feature extraction on the expanded image. Obtain the initial feature feats0 through the convolutional layer net1_conv0, as shown in formula (17). It has 4 input channels, 64 output channels, and a convolutional kernel size of 9.

[0154] feats0=net1_conv0(input_img)(17)

[0155] Step S140-3: Further deep extraction of preliminary features is performed through a combination of multiple convolutional layers and activation functions net1_convs to generate advanced features, as shown in formula (18).

[0156] featss=net1_convs(feats0)(18)

[0157] The structure of net1_convs consists of five stacks of convolutional layers and ReLU activation functions, with 64 input and output channels and a kernel size of 3.

[0158] Step S140-4: Reconstruct the high-level features by applying the convolutional layer net1_recon to convert the high-level features into the output channel image outs, as shown in formula (19). The input has 64 channels, the output has 4 channels, and the convolutional kernel size is 3.

[0159] outs = net1_recon(featss)(19)

[0160] Step S140-5: Normalize the output image using sigmoid to generate the reflection component R and the illumination component L. The reflection image contains color information of three channels, while the illumination image contains light intensity information of a single channel, as shown in formulas (20) and (21).

[0161] R=sigmoid(outs[:,0:3,∶,∶])(20)

[0162] L=sigmoid(outs[:,3:4,∶,∶])(21)

[0163] Through the above steps, the input image is decomposed into a reflection image R and an illumination image L. This network can effectively extract illumination and reflection information from the input image, providing rich feature data for subsequent image enhancement or restoration tasks.

[0164] Therefore, the reflection component and illumination component obtained from image I1 are R1 and L1; the reflection component and illumination component obtained from image I2 are R2 and L2.

[0165] Step S150: After swapping the reflection components of image pair 1, a new image pair 2 (I1′, I2′) is resynthesized.

[0166] As shown in the Interchange section of Figure 5, this embodiment specifically includes:

[0167] Step S150-1: Exchange R1 and R2, which are decomposed from images I1 and I2, as shown in formula (22).

[0168] (R1,L1),(R2,L2)→(R2,L1),(R1,L2)(22)

[0169] Step S150-2: The swapped R and L components are synthesized again to form a new image, as shown in formulas (23) and (24).

[0170] I′1=R2·L1(23)

[0171] I′2=R1·L2(24)

[0172] After exchanging the reflection components, they are re-fused into a new image, providing specific image data for subsequent image decomposition.

[0173] Step S160: Input image pair 2 into the feature encoder, and then decode it through the transposed convolutional decoder of the reflection layer to obtain R′1 and R′2;

[0174] like Figure 2 As shown below, the Feature Encoder and TransConvRefDecoder are specifically described in this embodiment:

[0175] Step S160-1: Input the re-synthesized image into the Feature Encoder for encoding, as shown in formula (25). The implementation of the Feature Encoder is the same as that described in step S120.

[0176] F i v = FeatureEncoder(I i ′)(25)

[0177] In step S160-2, the encoded features are input into TransConvRefDecoder for decoding, as shown in formula (26). The implementation of TransConvRefDecoder is the same as that described in step S130.

[0178] R′ i =TransConvRefDecoder(F i ′)(26)

[0179] After this encoding and decoding, the results are R′1 and R′2.

[0180] Step S170: Calculate various loss functions based on the decoding results to improve the accuracy of the re-identification network, including decoding loss. Reconstruction loss Reflector loss Triple loss and cross-entropy loss

[0181] like Figure 2 As shown, in this embodiment:

[0182] Step S170-1: Based on the series of results obtained above, calculate their loss functions respectively, as shown in formulas (27)(28)(29)(30)(31).

[0183]

[0184] exist In this context, MSE represents mean squared error, and SSIM represents the structural similarity index measure.

[0185] exist In this context, MAE represents the mean absolute error, and (R,R′) represents (R1,R′2) or (R2,R′1), respectively.

[0186] During network training, the performance of the model can be significantly improved by limiting and gradually reducing the magnitude of the calculated loss function value.

[0187] The above description is only for the purpose of helping to understand the method and core essence of the present invention, but the scope of protection of the present invention is not limited thereto. For those skilled in the art, any equivalent substitutions or modifications made to the technical solution and inventive concept disclosed in the present invention within the scope of the technology disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for pedestrian re-identification in low-light environments with zero samples, characterized in that, Includes the following steps: S1. A dark-light image synthesis network is used to process well-lit images to generate dark-light images with missing lighting. The well-lit images and the dark-light images with missing lighting are combined to obtain image pair 1. , ; S2. Extract features from image pair 1 using a feature encoder; S3. The features are processed by a transposed convolutional decoder through a reflection layer to obtain the decoding result. , Furthermore, the features are processed through a re-identification network to obtain the re-identification result; The transposed convolutional decoder of the reflection layer processes the data as follows: first, input Conv3 and activate it through the ReLU function, then input the output of the convolutional layer Conv3 into the transposed convolutional layer TransConv and activate it through the Sigmoid function; the transposed convolutional layer TransConv is a layer that can learn the optimal upsampling method and optimize the weights through backpropagation to adapt to specific data features. S4. Use the Retinex decomposition network to decompose the image in image pair 1 into reflection components. , and light component , ; S5. Swap the reflection components of image pair 1, and then resynthesize the swapped reflection components with the original illumination components to obtain image pair 2. S6. Image pair 2 is processed by a feature encoder to obtain features, and the features are then decoded by a transposed convolutional decoder to obtain the decoding result. , ; S7. Calculate the loss function based on the decoding and re-identification results of image pair 1 and image pair 2. The loss function includes the decoding loss. Reconstruction loss loss of reflective layer Triple loss and cross-entropy loss ; The loss function is as follows: Among them, In this context, MSE represents mean squared error, and SSIM represents the structural similarity index measure. exist In this context, MAE represents the mean absolute error. They represent ( , )or( , F represents the re-identification result; S8. Optimize the re-identification network by adjusting the network parameters according to the loss function.

2. The zero-sample pedestrian re-identification method in low-light environments according to claim 1, characterized in that, Image pair 1 is obtained in S1 ( , ), as detailed below: in, For well-lit images, DarkISP is a network for synthesizing dark-light images, which are characterized by a lack of illumination. The dark light image synthesis network generates dark light images with missing illumination, as follows: The well-lit image is processed sequentially with inverse tone mapping, inverse gamma correction, conversion from the standard sRGB color space to the camera's RGB color space, inverse white balance adjustment, brightness adjustment, noise reduction, white balance adjustment again, conversion from the camera's RGB color space back to the standard sRGB color space, and gamma correction to obtain a low-light image with low illumination characteristics.

3. The zero-sample pedestrian re-identification method in low-light environments according to claim 2, characterized in that, S2 is specifically as follows: Among them, Feature Encoder is the feature encoder. For image pair 1 ( , ), The features output by the feature encoder.

4. The zero-sample pedestrian re-identification method in low-light environments according to claim 3, characterized in that, S3 is specifically as follows: Among them, TransConvRefDecoder is a transposed convolutional decoder of the reflection layer, and ReIDNet is a re-identification network; The decoding result of image pair 1 , F represents the re-identification result.

5. The zero-sample pedestrian re-identification method in low-light environments according to claim 4, characterized in that, The transposed convolutional decoder of the reflection layer processes the data to obtain the decoding result, as follows: The encoder's encoded result is input into Conv3 and activated by the ReLU function, as shown in the following equation; Where Feat is the output of the feature encoder; The output DF1 of the convolutional layer Conv3 is input into the transposed convolutional layer TransConv and activated by the Sigmoid function, as shown in the following equation; in, The decoding result obtained after decoding, i.e., the reflection component. , .

6. The zero-sample pedestrian re-identification method in low-light environments according to claim 4, characterized in that, In step S4, the Retinex decomposition network is used to decompose image pair 1. Decomposed into , Similarly, images Decomposed into , The details are as follows: Among them, Retinex DecomNet is the Retinex decomposition network. The reflection component carries reflection information. The illumination component contains illuminance information; The Retinex decomposition network performs image decomposition as follows: The maximum value channel expansion process is performed on the input image Img to extract the maximum value channel input_max of the input image, and then it is concatenated with the original image to generate an image input_img containing more information; Shallow feature extraction is performed on the expanded image, and preliminary features are obtained through the convolutional layer net1_conv0. The preliminary features are then extracted in depth through the combination of multiple convolutional layers and activation functions net1_convs to generate high-level features. The high-level features are reconstructed by applying a convolutional layer net1_recon to convert them into the output channel image. The output channel image is then normalized using a sigmoid function to generate the reflection component R and the illumination component L.

7. The zero-sample pedestrian re-identification method in low-light environments according to claim 6, characterized in that, Image pair 2 is obtained in S5 ( , ), as detailed below: 。 8. The zero-sample pedestrian re-identification method in low-light environments according to claim 7, characterized in that, The decoding result is obtained in S6. , The details are as follows: in, For image pair 2 ( , ), The features output by the feature encoder, The decoding result of image pair 2 , .

9. A pedestrian re-identification system for low-light environments with zero samples, applicable to the method described in any one of claims 1-8, characterized in that, include: A low-light image synthesis network is used to generate low-light images with missing lighting. The feature encoder consists of two convolutional layers used to encode image features; A transposed convolutional decoder with a reflective layer is used to decode encoded features; The Retinex decomposition network is used to decompose image pairs into illumination and reflection components. The pedestrian re-identification network is a re-identification network with VGG16 network as its backbone, used to obtain pedestrian re-identification results in images.

Images