Low-light target detection method based on unsupervised image enhancement

Abstract

The invention provides a low-light target detection method based on unsupervised image enhancement. The low-light target detection method comprises the following steps: acquiring a to-be-adjusted image; constructing a low-light target detection data set based on unsupervised image enhancement; establishing an illumination curve adjustment network, and inputting the to-be-adjusted image into the illumination curve adjustment network to obtain a pre-intensified image after brightness correction; building an unsupervised image quality recovery network, and inputting the pre-enhanced image into the illumination curve adjustment network to obtain a final enhanced image; and using a target detection network to evaluate and test the enhanced image. According to the method, natural image prior of a pre-training diffusion model and a feature consistency learning mode based on target detection task guidance are combined, and feature representation more suitable for a target detection task is generated through a joint training mode while it is ensured that the visual quality of the image is remarkably improved; therefore, efficient and accurate detection of targets in night driving, automatic driving and low-light monitoring scenes is realized.

Description

Technical Field

[0001] This invention relates to the technical field of digital image enhancement and target detection, and in particular to a low-light target detection method based on unsupervised image enhancement. Background Technology

[0002] With the rapid development of technologies such as autonomous driving, intelligent transportation, and nighttime security monitoring, higher demands are being placed on target detection in low-light environments. However, traditional target detection algorithms often suffer from reduced recognition rates, increased false positives, and missed positives in low-light, noisy nighttime scenes. To improve target detection performance in low-light environments, researchers have attempted to enhance images before detection; however, simply increasing brightness may over-amplify noise or distorted textures, thus affecting detection accuracy. Recently emerging image enhancement methods based on diffusion models utilize prior distributions pre-trained on large-scale natural image datasets to perform more refined denoising and detail restoration on complex low-light degradation. However, without specific consideration for the detection task, the enhanced image may still be detrimental to subsequent target detection. Figure 1 As shown, some state-of-the-art low-light enhancement methods exhibit performance degradation in complex lighting scenarios with interference from streetlights, signs, and other light sources, particularly in downstream target detection tasks. Therefore, achieving a balance between image enhancement and target detection requirements within an unsupervised framework, while remaining effective in low-light environments, has become a pressing technical challenge.

[0003] Chinese invention patent application CN118447265A discloses a target detection method under low-light illumination conditions based on self-calibrated illumination learning, including the following steps: Step A, predicting illumination of the low-light image using a CNN network; Step B, constructing the CNN network using convolutional modules during Step A; Step C, learning the input of each stage and the initial stage using the CNN network; Step D, extracting features from the input image; Step E, performing regression detection based on feature maps of different scales using the extracted feature maps. The drawback of this method is that, due to the lack of an end-to-end joint training strategy or a detection-oriented feature alignment mechanism, the distribution of the enhanced image is difficult to fully match with the feature space distribution of the detector, easily causing distribution shift and affecting detection performance. Furthermore, the enhancement stage of this method does not introduce prior constraints from relevant natural images, and under its unsupervised setting, it is prone to problems such as color distortion and insufficient cross-domain generalization, further affecting subsequent detection performance and robustness. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a low-light target detection method based on unsupervised image enhancement. This method combines the natural image priors of a pre-trained diffusion model with a feature consistency learning approach oriented towards the target detection task. Through joint training, it significantly improves image visual quality while generating feature representations more suitable for target detection tasks, thereby achieving efficient and accurate target detection in nighttime driving, autonomous driving, and low-light monitoring scenarios.

[0005] The purpose of this invention is to provide a low-light target detection method based on unsupervised image enhancement, which includes acquiring an image to be adjusted and further includes the following steps:

[0006] Step 1: Construct a low-light target detection dataset based on unsupervised image enhancement;

[0007] Step 2: Construct an illumination curve adjustment network, input the image to be adjusted into the illumination curve adjustment network, and obtain a pre-enhanced image after brightness correction;

[0008] Step 3: Construct an unsupervised image quality restoration network, input the pre-enhanced image into the illumination curve adjustment network, and obtain the final enhanced image;

[0009] Step 4: Use an object detection network to test and evaluate the enhanced image.

[0010] Preferably, step 2 includes the following sub-steps:

[0011] Step 21: Label the low-light images in the low-light target detection dataset as I(x), and use them as input to the illumination curve adjustment network;

[0012] Step 22: Estimate the corresponding curve parameter α pixel by pixel to obtain the parameter estimation map A of I(x);

[0013] Step 23: Repeat step 22 until the preset number of iterations n is reached, so as to achieve progressive pixel-by-pixel exposure correction;

[0014] Step 24: Update parameter estimation plot A.

[0015] In any of the above schemes, step 24 preferably further includes adjusting the image again, and the intermediate enhancement result I n+1 The formula for (x) is:

[0016] I n+1 (x)= I n (x)+ A n (x) I n (x)(1- I n (x))

[0017] Among them, A n (x) is the pixel-level curve parameter estimation map corresponding to the nth iteration.

[0018] In any of the above schemes, preferably, the illumination curve adjustment network consists of six stacked convolutional layers containing skip connections. The first five layers consist of 3×3 convolutional kernels, batch normalization, and ReLU activation functions. The last layer uses the Tanh activation function to control the output curve parameters within the range of [-1, 1], thereby achieving pixel-by-pixel dynamic range adjustment of the illumination of the input image. The formula is as follows:

[0019]

[0020] Where x is the original low-light input image, This is the pre-enhanced image after curve adjustment, and curve(.) is the illumination curve adjustment network.

[0021] Preferably, in any of the above schemes, the unsupervised image quality restoration network utilizes a pre-trained diffusion model as its basic framework and employs a three-step sampling inverse denoising method based on the IDDPM sampling strategy to restore the quality of the pre-enhanced image. The working method includes:

[0022] Step 31: Build a YOLOv8 object detection network and connect the image quality restoration network to the object detection network for joint training;

[0023] Step 32: Utilize the Feature Consistency Learning (AFM) module guided by detection loss;

[0024] Step 33: Perform low-frequency enhancement on the features output by the backbone network of the diffusion model;

[0025] Step 34: Use the Sobel operator to extract edge information of shallow high-frequency features in skip connections to obtain the final enhanced image.

[0026] In any of the above schemes, it is preferred that the sampling process of the unsupervised image quality restoration network is guided by a detection loss, expressed as follows:

[0027]

[0028] Where, x t As the initial state of the diffusion model, α t and These are parameters for noise control during the diffusion process. For random noise, σ t For the noise scale in the corresponding step, This is a feature consistency learning module based on detection loss guidance.

[0029] In any of the above schemes, the preferred embodiment is that the AFM module samples the last three layers of the U-Net network at each step of the diffusion model, consisting of two parts: deep feature adaptive low-pass filtering and shallow feature high-frequency compensation.

[0030] In any of the above schemes, it is preferred that the deep backbone network features X are used when sampling the diffusion model in the AFM module. bone An adaptive low-pass filter is applied, combining local and global information to generate low-pass filter weights for each spatial location and channel. The filter calculation formula is as follows:

[0031]

[0032] Among them, Y low Let (c, i, j) be the low-pass filter weights, (c, i, j) be the channel, height, and width indices of the feature map, respectively, and S be the set of spatial locations centered at (i, j). The filter weights are generated for each spatial location and channel, where p is the vertical displacement index relative to position (i,j) within set S, q is the horizontal displacement index relative to position (i,j), and X... c,i+p,j+q The eigenvalues ​​of the backbone feature map at channel c and spatial location (i+p,j+q).

[0033] In any of the above schemes, the preferred option is that the filter weights generated for each spatial location and channel are... Defined as

[0034]

[0035] in, T represents the weight values ​​generated from local features. c These are the adjustment parameters predicted from global features.

[0036] Preferably, in any of the above schemes, the shallow features X from skip connections are processed in the AFM module. skip Useful high-frequency information is extracted using the Sobel operator and further enhanced by densely connected convolutions, as shown in the formula:

[0037] Y high =Conv(Sobel(X skip ))

[0038] Sobel(·) is used to perform edge detection operations, and Conv(·) is used to further enhance the extracted features.

[0039] In any of the above schemes, it is preferred that, after obtaining Y after adaptive low-pass filtering... low Y enhanced by Sobel operator highAfterwards, regarding Y low Upsampling is performed to adjust the feature size, followed by channel concatenation and 1×1 convolution to fuse the two feature components. Finally, the channel attention mechanism CA(·) is used to enhance the focus on useful features. The formula is as follows:

[0040] Y final =CA(Conv 1×1 ([Up(Y low ), Y high ]))

[0041] Up(·) is used for the spatial upsampling operator.

[0042] This invention proposes a low-light target detection method based on unsupervised image enhancement. It effectively bridges the interdomain differences between training data and real low-light driving scene data by utilizing the natural image prior provided by the pre-trained diffusion model. While achieving high-quality image enhancement, it significantly improves the accuracy and robustness of target detection in low-light autonomous driving scenarios.

[0043] IDDPM, Improved Denoising Diffusion Probabilistic Models, is an improved diffusion probability model. Attached Figure Description

[0044] Figure 1 This is a flowchart of a preferred embodiment of a low-light target detection method based on unsupervised image enhancement according to the present invention.

[0045] Figure 2 This is a schematic diagram of an embodiment of a CFLDP network based on an unsupervised image enhancement method for low-light target detection according to the present invention.

[0046] Figure 3 This is a schematic diagram of the structure of an embodiment of an illumination curve adjustment network for low-light target detection based on unsupervised image enhancement according to the present invention.

[0047] Figure 4 This is a flowchart of another preferred embodiment of a low-light target detection method based on unsupervised image enhancement according to the present invention.

[0048] Figure 5 This is an overall flowchart of an embodiment of a low-light target detection method based on unsupervised image enhancement according to the present invention, including dataset construction, preprocessing, joint training, and inference output.

[0049] Figure 6 This is a comparative schematic diagram of the target detection results of an embodiment of a low-light target detection method based on unsupervised image enhancement according to the present invention.

[0050] Figure 7 This is a comparative schematic diagram of the image enhancement results of an embodiment of a low-light target detection method based on unsupervised image enhancement according to the present invention. Detailed Implementation

[0051] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0052] Example 1

[0053] like Figure 1 As shown, a low-light target detection method based on unsupervised image enhancement is implemented in step 100, which involves acquiring the image to be adjusted.

[0054] Perform step 110 to construct a low-light target detection dataset based on unsupervised image enhancement;

[0055] Step 120 involves constructing an illumination curve adjustment network and inputting the image to be adjusted into the illumination curve adjustment network to obtain a pre-enhanced image after brightness correction. This includes the following sub-steps:

[0056] Execute step 121, label the low-light images in the low-light target detection dataset as I(x), and use them as input to the illumination curve adjustment network. The illumination curve adjustment network consists of 6 stacked convolutional layers with skip connections. The first five layers consist of 3×3 convolutional kernels, batch normalization, and ReLU activation functions. The last layer uses the Tanh activation function to control the output curve parameters within the range of [-1, 1], thereby achieving pixel-by-pixel dynamic range adjustment of the illumination of the input image. The formula is as follows:

[0057]

[0058] Where x is the original low-light input image, This is the pre-enhanced image after curve adjustment, and curve(.) is the illumination curve adjustment network.

[0059] Execute step 122 to estimate the corresponding curve parameter α pixel by pixel, and obtain the parameter estimation map A of I(x);

[0060] Execute step 123 and repeat step 22 until the preset number of iterations n (n=8) is reached, so as to achieve progressive pixel-by-pixel exposure correction;

[0061] Perform step 124, update parameter estimation map A, readjust the image, and enhance intermediate result I. n+1 The formula for (x) is:

[0062] I n+1 (x)= I n(x)+ A n (x) I n (x)(1- I n (x))

[0063] Among them, A n (x) is the pixel-level curve parameter estimation map corresponding to the nth iteration, which is used to adjust the exposure gain of this iteration.

[0064] Step 130 involves constructing an unsupervised image quality restoration network. The pre-enhanced image is input into the illumination curve adjustment network to obtain the final enhanced image. The unsupervised image quality restoration network utilizes a pre-trained diffusion model as its basic framework and employs a three-step sampling inverse denoising method based on the IDDPM sampling strategy to restore the quality of the pre-enhanced image. The working method includes:

[0065] Execute step 131 to build a YOLOv8 object detection network, and connect the image quality restoration network to the object detection network for joint training; the sampling process of the unsupervised image quality restoration network uses detection loss to guide the sampling process, expressed by the following formula:

[0066]

[0067] Where, x t As the initial state of the diffusion model, α t and These are parameters for noise control during the diffusion process. For random noise, σ t For the noise scale in the corresponding step, This is a feature consistency learning module based on detection loss guidance.

[0068] Step 132 is executed, utilizing the Feature Consistency Learning (AFM) module guided by detection loss. The AFM module samples the last three layers of the U-Net network at each step of the diffusion model, consisting of two parts: deep feature adaptive low-pass filtering and shallow feature high-frequency compensation.

[0069] The deep backbone network features X sampled during diffusion model sampling in the AFM module bone An adaptive low-pass filter is applied, combining local and global information to generate low-pass filter weights for each spatial location and channel. The filter calculation formula is as follows:

[0070]

[0071] Among them, Y low Let (c, i, j) be the low-pass filter weights, (c, i, j) be the channel, height, and width indices of the feature map, respectively, and S be the set of spatial locations centered at (i, j). The filter weights are generated for each spatial location and channel, where p is the vertical displacement index relative to position (i,j) within set S, q is the horizontal displacement index relative to position (i,j), and X... c,i+p,j+q The eigenvalues ​​of the backbone feature map at channel c and spatial location (i+p,j+q).

[0072] The filter weights generated for each spatial location and channel Defined as

[0073]

[0074] in, T represents the weight values ​​generated from local features. c These are the adjustment parameters predicted from global features.

[0075] In the AFM module, shallow features X from skip connections skip Useful high-frequency information is extracted using the Sobel operator and further enhanced by densely connected convolutions, as shown in the formula:

[0076] Y high =Conv(Sobel(X skip ))

[0077] Sobel(·) is used to perform edge detection operations, and Conv(·) is used to further enhance the extracted features.

[0078] After obtaining Y after adaptive low-pass filtering, low Y enhanced by Sobel operator high Afterwards, regarding Y low Upsampling is performed to adjust the feature size, followed by channel concatenation and 1×1 convolution to fuse the two feature components. Finally, the channel attention mechanism CA(·) is used to enhance the focus on useful features. The formula is as follows:

[0079] Y final =CA(Conv 1×1 ([Up(Y low ), Y high ]))

[0080] Where Up(·) is the spatial upsampling operator, used to upsample Y. low Upsampling to Y high Same spatial resolution to achieve scale alignment.

[0081] Step 133 is executed to perform low-frequency enhancement on the features output by the backbone network of the diffusion model.

[0082] Step 134 is executed, using the Sobel operator to extract edge information of shallow high-frequency features in skip connections, to obtain the final enhanced image.

[0083] Perform step 140, using an object detection network to test and evaluate the enhanced image.

[0084] Example 2

[0085] To address the problems of severe image noise, loss of detail, and low target detection accuracy caused by low-light environments in existing low-light target detection methods, this invention provides a low-light target detection method based on unsupervised image enhancement, CFLDP, such as... Figure 2 As shown. The proposed method combines the natural image prior of a pre-trained diffusion model with a feature consistency learning approach guided by the target detection task. Through joint training, it generates feature representations more suitable for target detection tasks while significantly improving image visual quality, thereby achieving efficient and accurate target detection in nighttime driving, autonomous driving, and low-light monitoring scenarios. The CFLDP method consists of two main parts: a zero-reference illumination curve adjustment network and an image quality restoration network based on detection loss-guided feature consistency learning and diffusion prior.

[0086] 1. Zero-reference illumination curve adjustment network

[0087] The illumination curve adjustment network employs a structure consisting of six stacked convolutional layers containing skip connections, such as... Figure 3 As shown in the diagram, the first five layers of the network consist of 3×3 convolutional kernels, batch normalization, and ReLU activation functions. The last layer uses the Tanh activation function to control the output curve parameters within the range of [-1, 1], thereby achieving pixel-by-pixel dynamic range adjustment of the input image's illumination, as shown in the following formula:

[0088]

[0089] Where x represents the original low-light input image, This represents the pre-enlarged image after curve adjustment. After initial illumination correction using curve functions, The brightness and contrast are closer to the distribution of normal light images, providing a more balanced initial state for subsequent image quality restoration.

[0090] 2. Image Quality Restoration Network Based on Feature Consistency Learning Guided by Detection Loss and Diffusion Prior

[0091] The image quality restoration network based on detection loss-guided feature consistency learning and diffusion prior is designed to utilize a diffusion model pre-trained on large-scale natural image datasets such as ImageNet. Through a three-step backsampling process based on the IDDPM sampling strategy, it denoises and restores details in the pre-enhanced image, making the enhanced image distribution closer to that of a real natural image and effectively bridging the inter-domain differences between training data and real low-light driving scenarios. Furthermore, during the sampling process, a detection loss-guided feature consistency learning (AFM) module is designed and used to enhance the low-frequency features required for downstream object detection tasks while preserving high-frequency texture and edge information, further restoring image quality.

[0092] In the network, the pre-enhanced image is considered the initial state of the diffusion model. During sampling, the inherent noise and artifacts in the image are gradually removed, while the structural information and texture details obscured by low-light conditions in the original image are recovered. The sampling process is shown in the following formula:

[0093]

[0094] Where, x t As the initial state of the diffusion model, α t and These are parameters for noise control during the diffusion process. It is random noise, σ t This refers to the noise scale at the corresponding step. (Function) It is based on a feature consistency learning module guided by detection loss, and the specific implementation idea is as follows.

[0095] The AFM module operates on the last three layers of the U-Net network at each step of the diffusion model, and is divided into two parts, as follows: Figure 4 As shown, one part is the deep feature adaptive low-pass filtering part, and the other part is the shallow feature high-frequency compensation part.

[0096] Deep features of the backbone network during diffusion model sampling Despite containing global semantic information, high-frequency interference from noise and other internal inconsistencies can negatively impact the detection task. To address this issue, we reference feature consistency learning and incorporate it into the AFM module. An adaptive low-pass filter is applied, combining local and global information to generate low-pass filter weights for each spatial location and channel, modulating changes in signal frequency to improve intra-class consistency of features. The filter calculation formula is as follows:

[0097]

[0098] Where (c,i,j) represent the channel, height, and width indices of the feature map, respectively, and S represents the set of spatial locations centered at (i,j). These are the filter weights generated for each spatial location and channel, defined as shown in the following formula:

[0099]

[0100] in, T represents the weight values ​​generated from local features. c The adjustment parameters, predicted from global features, control the smoothness of the filter. By combining local features V and global features T, this part dynamically adjusts the filter response to enhance low-frequency signals while suppressing high-frequency noise. Furthermore, Softmax normalization ensures the low-pass properties of the weights, effectively preserving both global and local structural information of the target.

[0101] To further restore image quality and extract edge and texture information from high-frequency components that are heavily contaminated by noise, we process shallow features X from skip connections in the AFM module. skip Use the Sobel operator to extract useful high-frequency information, which is further enhanced by densely connected convolutions, as shown in the following formula:

[0102] Y high =Conv(Sobel(X skip ))

[0103] Sobel(·) is used to perform edge detection operations, and Conv(·) is used to further enhance the extracted features. Let Y... high It can effectively compensate for the texture loss caused by adaptive low-pass filtering in the backbone network.

[0104] After obtaining Y after adaptive low-pass filtering, low Y enhanced by Sobel operator high Afterwards, for Y low Upsampling is performed to adjust the feature size, followed by channel concatenation and 1x1 convolution to fuse the two features. Finally, the channel attention mechanism CA(·) is used to enhance the attention to useful features, as shown in the following formula.

[0105] Y final =CA(Conv 1×1 ([Up(Y low ), Y high ]))

[0106] During the training of the network, in order to improve the adaptability of the generated images in the downstream object detection task, a joint training method with the YOLOv8-based object detection network is used, and the detection loss is used as a supervision signal to guide the learning of the AFM module.

[0107] The task-oriented training approach described above ensures that the image restoration network not only relies on the noise scheduling parameters pre-trained by the diffusion model, but also adjusts based on the error feedback from the object detection network after detecting the current image. Specifically, the AFM module makes the noise prediction function more inclined to generate images with detection-friendly features. This approach effectively suppresses high-frequency noise detrimental to the detection task while fully enhancing low-frequency structural information, and retains useful high-frequency texture and edge information, thus balancing image enhancement quality with the performance of downstream tasks.

[0108] During the training and fine-tuning process, the image restoration network, except for the AFM module, and the object detection network are in a frozen state. The network can usually achieve the expected results after 10 rounds of fine-tuning.

[0109] Compared with the prior art, the present invention has the following advantages and technical effects:

[0110] The method proposed in this invention first pre-enhances the low-light image by adjusting the illumination curve, and then uses a pre-trained diffusion model to construct an image quality restoration network to further remove noise and artifacts from the low-light image, restoring the texture details and edge information of the image. Unlike traditional image enhancement algorithms that only focus on a single image enhancement, this invention introduces the detection loss into the sampling process of the diffusion model through joint training. Specifically, using the Feature Consistency Learning (AFM) module guided by the detection loss, the deep features of the diffusion model backbone network are low-pass filtered, and the shallow high-frequency features in skip connections are edge-extracted using the Sobel operator. While preserving the necessary image texture and edge information, this method strengthens the low-frequency features required for target detection, significantly improving the discrimination ability of the detection network in complex low-light scenes.

[0111] Furthermore, the joint training strategy adopted in this invention only requires 10 rounds of fine-tuning to achieve the expected results, and the training does not rely on paired data, which greatly reduces the training cost. It has good accuracy and robustness in target detection tasks in different low-light driving scenarios, providing a stable and low-cost vision solution for fields such as autonomous driving and night monitoring.

[0112] Example 3

[0113] This invention relates to the fields of digital image processing and object detection, and discloses a low-light object detection method based on unsupervised image enhancement. Addressing the challenge of object detection in low-light autonomous driving scenarios, this invention designs an image enhancement-based object detection method, comprising the following steps: constructing a low-light image enhancement dataset; building an illumination curve adjustment network to perform preliminary brightness correction on the low-light images, obtaining pre-enhanced images; constructing an unsupervised image quality restoration network and an object detection network and jointly training them, embedding the pre-enhanced images into the sampling process of a pre-trained diffusion model to restore image quality and obtain enhanced images that conform to the distribution of the real image; simultaneously introducing an adaptive frequency modulation module guided by detection loss during the sampling process to further guide the enhanced images to generate the low-frequency features required for object detection and suppress high-frequency noise interference; and testing using the object detection network to obtain object detection results. This invention does not require paired training data, and the joint training phase requires only a very small number of training rounds (10 rounds) to achieve the desired effect. The proposed method effectively bridges the interdomain differences between training data and real low-light driving scene data by utilizing the natural image prior provided by the pre-trained diffusion model. While achieving high-quality image enhancement, it significantly improves the accuracy and robustness of target detection in low-light autonomous driving scenarios.

[0114] like Figure 5 As shown, a low-light target detection method based on unsupervised image enhancement includes the following steps:

[0115] Step S1: Construct a low-light target detection dataset based on unsupervised image enhancement. The specific process is as follows:

[0116] Step S11: Select low-light images of various scenes from existing public low-light target detection datasets;

[0117] Step S12: Collect and label real nighttime driving scene datasets, remove some blurry and overexposed images, and uniformly adjust the images to (1080, 720); the dataset contains 3000 low-light RGB images;

[0118] Step S13: Merge the above-mentioned public low-light target detection dataset and the acquired images to obtain a low-light target detection dataset.

[0119] Step S2: Construct an illumination curve adjustment network consisting of six stacked convolutional layers with skip connections to estimate curve parameters. Each preceding convolutional layer uses the ReLU activation function, and the last layer uses the Tanh activation function to keep the curve parameters between [-1, 1]. Label the low-light images in the low-light target detection dataset from Step S1 as I(x), and use I(x) as the input to the illumination curve adjustment network to estimate the corresponding curve parameters α pixel by pixel, obtaining the parameter estimation map A of I(x), which is used to adjust the dynamic range of I(x). Next, apply the network again to update the parameter estimation map and adjust the image again, enhancing the intermediate result I. n+1 The formula for representing (x) is shown below:

[0120] I n+1 (x)= I n (x)+ A n (x) I n (x)(1- I n (x))

[0121] Next, the above network is applied iteratively. Throughout the process, a set of unsupervised loss constraints are used to learn and obtain the optimally fitted curve parameter estimation map A, and finally, the pre-enhanced image after brightness correction is obtained.

[0122] Step S3: Construct an unsupervised image quality restoration network. This network uses a pre-trained diffusion model as its basic framework and employs a three-step reverse denoising method based on the IDDPM sampling strategy to restore the quality of the pre-enhanced image, eliminate noise and artifacts, and restore structural information and texture details, so that the brightness, contrast and color distribution of the output image are closer to the real scene.

[0123] The unsupervised image quality restoration network is built on the IDDPM diffusion model pre-trained on the ImageNet large-scale natural image dataset. The model performs noise reduction and detail restoration on the pre-enhanced image through a three-step backsampling process to bridge the inter-domain differences between the training data and real low-light driving scenes. The sampling process is guided by a detection loss, expressed by the following formula:

[0124]

[0125] in, and These are parameters used to manage noise during the diffusion process, where z is random noise. This indicates the noise scale corresponding to this step. Function This represents the feature consistency learning module that guides the detection loss.

[0126] The AFM module performs adaptive low-pass filtering on the deep features extracted from the backbone network in the diffusion model. It enhances low-frequency information by generating low-pass filter weights based on the fusion of local features and global information. In addition, the AFM module extracts shallow features from skip connections, extracts edge information of the features using the Sobel edge extraction operator, and then obtains high-frequency texture information through two 1x1 convolutional layers and upsampling. Subsequently, the above low-frequency and high-frequency features are fused through the channel attention (CA) module.

[0127] Step S4: Build a YOLOv8 object detection network and pre-train it using a public dataset composed of normal light images; then connect the image quality restoration network to the object detection network for joint training; during the backsampling process of the pre-trained diffusion model described in Step S3, use the detection loss output of the object detection network to supervise the learning of the adaptive frequency modulation module, and correct the feature distribution of the image by fine-tuning the sampling process of the diffusion model, guiding the enhanced image to generate more low-frequency features that are beneficial to the detection network, while suppressing high-frequency noise, to obtain the final enhanced image; the adaptive frequency modulation module performs low-frequency enhancement on the features output by the backbone network of the diffusion model; and uses the Sobel operator to extract the edge information of shallow high-frequency features in the skip connections, further restoring the texture details of the image, while suppressing the high-frequency noise that is over-enhanced in the pre-enhancing stage.

[0128] Step S5: Use an object detection network to test and evaluate the enhanced images obtained in the above steps to verify the effectiveness of the method; at the same time, use image quality evaluation metrics such as NIQE, BRISQUE, and PIQE to evaluate the image enhancement effect of the proposed method.

[0129] The method adopts the YOLOv8 target detection network structure in the target detection part. This structure effectively improves the accuracy and real-time performance of target detection in low light environment through multi-scale feature fusion, dense connection and lightweight design, and further ensures accurate detection of vehicles, pedestrians and obstacles in low light autonomous driving scenarios.

[0130] The method is applicable to low-light driving scenarios and autonomous driving systems. It can effectively improve the accuracy and robustness of target detection in complex low-light environments, enhance the image to have good visual quality, and provide reliable visual support for actual driving safety and autonomous driving systems.

[0131] The method does not require paired training data and adopts unsupervised training. It achieves rapid enhancement of low-light images and simultaneous improvement of target detection performance through joint training with natural image priors provided by the pre-trained diffusion model and detection guidance.

[0132] Example 4

[0133] A low-light target detection method based on unsupervised image enhancement includes the following steps:

[0134] Step S1: Construct a low-light target detection dataset based on unsupervised image enhancement;

[0135] Step S2: Construct an illumination curve adjustment network consisting of six stacked convolutional layers with skip connections to estimate curve parameters. Each preceding convolutional layer uses the ReLU activation function, and the last layer uses the Tanh activation function to keep the curve parameters within the range of [-1, 1]. Label the low-light images in the low-light target detection dataset from Step S1 as I(x), and use I(x) as the input to the illumination curve adjustment network to estimate the corresponding curve parameters α pixel-by-pixel, obtaining the parameter estimation map A of I(x), which is used to adjust the dynamic range of I(x). Then, the network is applied iteratively to update the parameter estimation map A. Throughout the process, a set of unsupervised loss constraints is used to learn and obtain the pre-enhanced image after brightness correction.

[0136] Step S3: Construct an unsupervised image quality restoration network. This network uses a pre-trained diffusion model as its basic framework and employs a three-step sampling inverse denoising method based on the IDDPM sampling strategy to restore the quality of the pre-enhanced image, eliminating noise and artifacts. Construct a YOLOv8 object detection network and connect the image quality restoration network to the object detection network for joint training. Use the Feature Consistency Learning (AFM) module guided by the detection loss to perform low-frequency enhancement on the features output by the backbone network of the diffusion model. Finally, use the Sobel operator to extract the edge information of shallow high-frequency features in skip connections to obtain the final enhanced image.

[0137] Step S4: The object detection network is tested and evaluated on the enhanced image; at the same time, the image enhancement effect of the proposed method is evaluated using image quality evaluation metrics such as NIQE, BRISQUE, and PIQE.

[0138] like Figure 6 and Figure 7 As shown, (a) is the original image of the real scene, and (b) is the result image after image enhancement.

[0139] To better understand this invention, specific embodiments have been described in detail above, but these are not intended to limit the invention. Any simple modifications made to the above embodiments based on the technical essence of this invention still fall within the scope of this invention. Each embodiment in this specification focuses on its differences from other embodiments; similar or identical parts between embodiments can be referred to mutually. For system embodiments, since they basically correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

Claims

1. A low-light target detection method based on unsupervised image enhancement, comprising acquiring an image to be adjusted, and further comprising the following steps: Step 1: Construct a low-light target detection dataset based on unsupervised image enhancement; Step 2: Construct an illumination curve adjustment network, input the image to be adjusted into the illumination curve adjustment network, and obtain a pre-enhanced image after brightness correction; Step 3: Construct an unsupervised image quality restoration network, input the pre-enhanced image into the illumination curve adjustment network, and obtain the final enhanced image; Step 4: Use an object detection network to test and evaluate the enhanced image.

2. The low-light target detection method based on unsupervised image enhancement as described in claim 1, characterized in that, Step 2 includes the following sub-steps: Step 21: Label the low-light images in the low-light target detection dataset as I(x), and use them as input to the illumination curve adjustment network; Step 22: Estimate the corresponding curve parameter α pixel by pixel to obtain the parameter estimation map A of I(x); Step 23: Repeat step 12 until the preset number of iterations n is reached, so as to achieve progressive pixel-by-pixel exposure correction; Step 24: Update parameter estimation plot A.

3. The low-light target detection method based on unsupervised image enhancement as described in claim 2, characterized in that, The illumination curve adjustment network consists of six stacked convolutional layers with skip connections. The first five layers are composed of 3×3 convolutional kernels, batch normalization, and ReLU activation functions. The last layer uses the Tanh activation function to control the output curve parameters within the range of [-1, 1], thereby achieving pixel-by-pixel dynamic range adjustment of the illumination of the input image. The formula is as follows: ， Where x is the original low-light input image, This is the pre-enhanced image after curve adjustment, and curve(.) is the illumination curve adjustment network.

4. The low-light target detection method based on unsupervised image enhancement as described in claim 3, characterized in that, The unsupervised image quality restoration network utilizes a pre-trained diffusion model as its basic framework and employs a three-step sampling inverse denoising method based on the IDDPM sampling strategy to restore the quality of the pre-enhanced image. The working method includes: Step 31: Build a YOLOv8 object detection network and connect the image quality restoration network to the object detection network for joint training; Step 32: Utilize the Feature Consistency Learning (AFM) module guided by detection loss; Step 33: Perform low-frequency enhancement on the features output by the backbone network of the diffusion model; Step 34: Use the Sobel operator to extract edge information of shallow high-frequency features in skip connections to obtain the final enhanced image.

5. The low-light target detection method based on unsupervised image enhancement as described in claim 4, characterized in that, The unsupervised image quality restoration network uses a detection loss to guide the sampling process, as expressed in the following formula: ， Where, x t As the initial state of the diffusion model, α t and These are parameters for noise control during the diffusion process. For random noise, σ t For the noise scale in the corresponding step, This is a feature consistency learning module based on detection loss guidance.

6. The low-light target detection method based on unsupervised image enhancement as described in claim 5, characterized in that, The AFM module samples the last three layers of the U-Net network at each step of the diffusion model, consisting of two parts: deep feature adaptive low-pass filtering and shallow feature high-frequency compensation.

7. The low-light target detection method based on unsupervised image enhancement as described in claim 6, characterized in that, The deep backbone network features X sampled during diffusion model sampling in the AFM module bone An adaptive low-pass filter is applied, combining local and global information to generate low-pass filter weights for each spatial location and channel. The filter calculation formula is as follows: ， Among them, Y low Let (c, i, j) be the low-pass filter weights, (c, i, j) be the channel, height, and width indices of the feature map, respectively, and S be the set of spatial locations centered at (i, j). The filter weights are generated for each spatial location and channel, where p is the vertical displacement index relative to position (i,j) within set S, q is the horizontal displacement index relative to position (i,j), and X... c,i+p,j+q The eigenvalues ​​of the backbone feature map at channel c and spatial location (i+p,j+q).

8. The low-light target detection method based on unsupervised image enhancement as described in claim 7, characterized in that, The filter weights generated for each spatial location and channel Defined as ， in, T represents the weight values ​​generated from local features. c These are the adjustment parameters predicted from global features.

9. The low-light target detection method based on unsupervised image enhancement as described in claim 8, characterized in that, In the AFM module, shallow features X from skip connections skip Useful high-frequency information is extracted using the Sobel operator and further enhanced by densely connected convolutions, as shown in the formula: Y high =Conv(Sobel(X skip )), Sobel(·) is used to perform edge detection operations, and Conv(·) is used to further enhance the extracted features.

10. The low-light target detection method based on unsupervised image enhancement as described in claim 8, characterized in that, After obtaining Y after adaptive low-pass filtering, low Y enhanced by Sobel operator high Afterwards, regarding Y low Upsampling is performed on its feature size. Subsequently, channel concatenation and 1×1 convolution are used to fuse the two parts of features. Finally, the channel attention mechanism CA(·) is used to enhance the attention to useful features. The formula is as follows: AND final =CA(Conv 1×1 ([Up(Y low ), AND high ])), Up(·) is used for the spatial upsampling operator.

Images