Image enhancement method, device, equipment, medium and product for target detection

Abstract

The application provides an image enhancement method, device, equipment, medium and product for target detection, relates to the technical field of image processing, and the method comprises the following steps: respectively performing image enhancement on a to-be-enhanced image to obtain a first enhanced image output by a supervised enhancement model and a second enhanced image output by an unsupervised enhancement model; fusing a luminance component of the first enhanced image and a luminance component of the second enhanced image to obtain a fused luminance component; and determining a target image for target detection based on the fused luminance component and a chroma component of the second enhanced image. The application decouples image enhancement from target detection, does not need to rely on an image pair and accurate labeling, is light, fuses supervised and unsupervised luminance components, can maintain image naturalness, avoids color deviation and the introduction of artifacts, improves semantic information of the target image, and is beneficial to improving target detection accuracy.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image enhancement method, apparatus, device, medium and product for target detection. Background Technology

[0002] Ambient lighting conditions are complex and variable, often exhibiting problems such as insufficient illumination, severe noise interference, and low contrast, leading to degraded image quality and consequently affecting the accuracy and stability of subsequent critical tasks such as target detection, recognition, and tracking. Currently, low-light image enhancement techniques mainly include traditional image processing methods, end-to-end supervised deep learning methods, and unsupervised / self-supervised deep learning methods.

[0003] Traditional image processing methods include histogram equalization, Retinex theory and its variants. These methods improve visual visibility by adjusting the brightness distribution of an image or separating the illumination and reflection components. They have advantages such as being computationally lightweight and requiring no training. However, they rely on manual parameter settings, and the enhanced results are prone to problems such as color distortion and blurred edges. Furthermore, the optimization targets are focused on human eye perception or general image quality indicators, without considering the adaptability to high-level semantic tasks such as object detection.

[0004] End-to-end supervised deep learning methods train image augmentation networks and object detection networks together, directly using detection performance metrics as optimization targets. While they can improve task accuracy on specific datasets, they are highly dependent on a large number of pairs of low-light and normal-light images and accurate annotations. Data acquisition is difficult, training costs are high, and the model structure is complex with low inference efficiency. They usually do not output intermediate augmented images, lack modular flexibility, and are difficult to adapt to different detectors or reuse in other vision tasks.

[0005] Unsupervised / self-supervised deep learning methods achieve enhancement by designing self-supervised loss functions such as exposure control, color consistency, and spatial smoothing. They have good generalization ability, lightweight deployment, and image naturalness under unpaired data conditions, making them more suitable for real-world applications. However, due to the lack of real, normally lit images as distribution guides, their enhancement results often suffer from uneven illumination distribution, obvious color histogram spikes, and local contrast jumps, leading to blurred semantic structure and loss of texture details. This makes it difficult to effectively support high-level vision tasks with high requirements for semantic integrity, such as object detection.

[0006] Therefore, it is necessary to provide a low-light image enhancement solution that can maintain deployment portability and image naturalness while effectively improving target detection accuracy. Summary of the Invention

[0007] This application provides an image enhancement method, apparatus, device, medium, and product for target detection, which can maintain easy deployment and natural image quality, and effectively improve the accuracy of target detection.

[0008] This application provides an image enhancement method for target detection, including: Acquire the image to be enhanced, and determine a first enhanced image for supervised enhancement and a second enhanced image for unsupervised enhancement of the image to be enhanced; Extract the luminance component of the first enhanced image, and extract the luminance and chrominance components of the second enhanced image; The luminance components of the first enhanced image and the second enhanced image are fused to obtain a fused luminance component; Based on the fused luminance component and the chrominance component of the second enhanced image, a target image for target detection is determined.

[0009] According to the image enhancement method for target detection provided in this application, the step of fusing the luminance components of the first enhanced image and the second enhanced image to obtain a fused luminance component includes: The luminance components of the first enhanced image and the second enhanced image are weighted and fused based on a fusion network to obtain the fused luminance component. The fusion network includes a deep learning model for image fusion.

[0010] According to the image enhancement method for target detection provided in this application, the loss function of the fusion network includes a gradient loss function and a pixel intensity loss function.

[0011] According to the image enhancement method for target detection provided in this application, the step of weightedly fusing the luminance components of the first enhanced image and the second enhanced image based on a fusion network to obtain the fused luminance component includes: The first loss function weight and the second loss function weight of the fusion network are determined. The first loss function weight is used to determine the similarity between the luminance component output by the fusion network and the luminance component of the first enhanced image. The second loss function weight is used to determine the similarity between the luminance component output by the fusion network and the luminance component of the second enhanced image. The first loss function weight is greater than the second loss function weight. The fusion network performs weighted fusion of the luminance components of the first enhanced image and the second enhanced image based on the weights of the first loss function and the second loss function to obtain the fused luminance component.

[0012] According to the image enhancement method for target detection provided in this application, before extracting the luminance component of the first enhanced image and extracting the luminance and chrominance components of the second enhanced image, the method further includes: Perform a first color space conversion on the first enhanced image and the second enhanced image; The step of determining the target image for target detection based on the fused luminance component and the chrominance component of the second enhanced image includes: A second color space conversion is performed on the fused luminance component and the chrominance component of the second enhanced image to obtain the target image for target detection; The first color space conversion is related to the second color space conversion.

[0013] According to the image enhancement method for target detection provided in this application, it further includes: The target image is input into any preset target detector to obtain the target detection result output by the target detector.

[0014] This application also provides an image enhancement apparatus for target detection, comprising: The first enhancement module is used to acquire the image to be enhanced and determine a first enhanced image and a second enhanced image for supervised enhancement of the image to be enhanced; An extraction module is used to extract the luminance component of the first enhanced image and the luminance and chrominance components of the second enhanced image. The fusion module is used to fuse the luminance components of the first enhanced image and the second enhanced image to obtain a fused luminance component. The second enhancement module is used to determine a target image for target detection based on the fused luminance component and the chrominance component of the second enhanced image.

[0015] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the image enhancement method for target detection as described above.

[0016] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the image enhancement method for target detection as described above.

[0017] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the image enhancement method for target detection as described above.

[0018] This application provides an image enhancement method, apparatus, device, medium, and product for target detection. The method involves acquiring an image to be enhanced, determining a first enhanced image to be enhanced under supervised supervision, and a second enhanced image to be enhanced under unsupervised supervision. The method extracts the luminance component of the first enhanced image and extracts the luminance and chrominance components of the second enhanced image. The luminance components of the first and second enhanced images are then fused to obtain a fused luminance component. Based on the fused luminance component and the chrominance component of the second enhanced image, a target image for target detection is determined. This application decouples image enhancement from target detection, eliminating the need for image pairs and precise annotation, thus offering portability. The fusion of supervised and unsupervised luminance components maintains image naturalness, avoids color shifts and artifact introduction, enhances the semantic information of the target image, and improves target detection accuracy. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is one of the flowcharts of the image enhancement method for target detection provided in this application.

[0021] Figure 2 This is a schematic diagram illustrating the differences in image components between the YCbCr color space and the RGB color space.

[0022] Figure 3 This is a schematic diagram of the structure of the fusion network provided in this application.

[0023] Figure 4 This is the second flowchart of the image enhancement method for target detection provided in this application.

[0024] Figure 5 This is a schematic diagram comparing the image enhancement effects of this application with existing algorithms.

[0025] Figure 6 This is a schematic diagram comparing the image enhancement effects of different fusion schemes for the chromaticity components of the second enhanced image.

[0026] Figure 7 This is a schematic diagram of the image enhancement device for target detection provided in this application.

[0027] Figure 8 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] This application focuses on task-oriented enhancement, which, under unsupervised conditions, utilizes a fusion of illumination component guidance mechanisms to enhance images. This not only ensures good visual naturalness but also effectively preserves or recovers semantic structures and texture features beneficial for object detectors, providing underlying support for building robust, efficient, and practical intelligent vision systems. This application can be applied to low-light intelligent scenarios that rely on visual perception and does not require acquiring large amounts of paired images and data annotations. It is suitable for fields with high requirements for object detection accuracy and system stability, such as autonomous driving nighttime perception scenarios, urban intelligent security monitoring scenarios, and industrial unmanned inspection robots.

[0030] The autonomous driving nighttime perception scenario specifically refers to the use of preprocessed images from vehicle-mounted cameras in low-light environments such as nighttime or tunnels, as described in this application, to improve the detection recall rate of pedestrians, vehicles, and traffic signs, thereby ensuring the safe operation of the autonomous driving system.

[0031] Urban intelligent security monitoring scenarios specifically refer to low-light monitoring scenarios such as nighttime streets, underground parking garages, and corridors, where enhanced video quality is input into the detection model to improve the accuracy of tasks such as abnormal behavior, suspicious targets, and license plate recognition.

[0032] In environments with insufficient lighting, such as factories, power stations, and mines, industrial unmanned inspection robots are equipped with vision sensors and use the images enhanced in real time through this application to ensure stable identification of targets such as equipment status, instrument readings, and safety signs.

[0033] The present application will now be described in detail with reference to the accompanying drawings.

[0034] Figure 1 This is one of the flowcharts illustrating the image enhancement method for object detection provided in this application, such as... Figure 1 As shown, this application provides an image enhancement method for target detection, which may include steps S100-S400.

[0035] Step S100: Obtain the image to be enhanced, and determine the first enhanced image to be enhanced under supervised supervision and the second enhanced image to be enhanced under unsupervised supervision.

[0036] Step S200: Extract the luminance component of the first enhanced image, and extract the luminance and chrominance components of the second enhanced image.

[0037] Step S300: The luminance components of the first enhanced image and the second enhanced image are fused to obtain a fused luminance component.

[0038] Step S400: Based on the fused luminance component and the chrominance component of the second enhanced image, determine the target image for target detection.

[0039] This application inputs the image to be enhanced into a supervised enhancement model and an unsupervised enhancement model, respectively. The supervised enhancement model performs supervised enhancement on the image to be enhanced to obtain a first enhanced image and extracts the luminance component of the first enhanced image. The unsupervised enhancement model performs unsupervised enhancement on the image to be enhanced to obtain a second enhanced image and extracts the luminance and chrominance components of the second enhanced image. Among them, the supervised enhancement model is a model pre-trained based on an existing low-light-normal-light image pair dataset, which only requires a small number of training samples for pre-training, without the need for a large number of image pairs and labeled data. Moreover, the supervised enhancement model is decoupled from the object detector, resulting in lower training costs.

[0040] The luminance component reflects the lightness and darkness of an image, while the chrominance component contains the image's color information. Methods for extracting the luminance component include converting the image to grayscale and directly extracting the Y channel in color spaces such as YUV / YCbCr. Methods for extracting the chrominance component include extracting the Cb and Cr channels (blue and red color differences) in YUV / YCbCr, extracting the H (hue) and S (saturation) channels in the HSV / HSL color space, and calculating the color difference in the RGB color space.

[0041] The methods for fusing the luminance components of the first enhanced image and the second enhanced image include simple pixel / region-based fusion, multi-scale decomposition-based fusion, sparse representation-based fusion, deep learning-based fusion, etc., and the embodiments of this application do not limit these methods.

[0042] The luminance component of the fused image and the chrominance component of the second enhanced image are combined and converted into an image to obtain a target image for target detection.

[0043] Understandably, this application decouples image enhancement from object detection, eliminating the need for image pairs and precise annotations. This lightweight approach, by fusing supervised and unsupervised brightness components, maintains image naturalness, avoids color shifts and artifacts, and enhances the semantic information of the target image, thus improving object detection accuracy. This application presents an unsupervised enhancement strategy that requires neither paired data nor joint training. By guiding the enhanced image to approximate the feature distribution under normal lighting, it significantly improves the accuracy of downstream detectors, avoiding the dependence of end-to-end methods on large amounts of labeled data.

[0044] As an optional embodiment, before extracting the luminance component of the first enhanced image and extracting the luminance and chrominance components of the second enhanced image, the method further includes: Perform a first color space conversion on the first enhanced image and the second enhanced image; The step of determining the target image for target detection based on the fused luminance component and the chrominance component of the second enhanced image includes: A second color space conversion is performed on the fused luminance component and the chrominance component of the second enhanced image to obtain the target image for target detection; The first color space conversion is related to the second color space conversion.

[0045] Commonly used image color spaces include RGB, HSV, YUV, and YCbCr. The RGB (Red, Green, Blue) color space is built based on the principle of human color recognition, representing most colors through different intensities of three primary colors. Although RGB is the most widely used color model in hardware devices such as monitors and color video cameras, it is less commonly used in scientific research because its structure makes it difficult to independently adjust hue, brightness, and saturation digitally. The YUV color space, on the other hand, allows brightness adjustment independent of color information. Each color consists of a luminance signal Y and two chrominance signals U (corresponding to the blue difference) and V (corresponding to the red difference). The Y signal directly reflects the luminance information in visual perception. YUV is derived from complete RGB color information. First, a black and white image is generated as the luminance channel, and then the two chrominance channels are extracted to reduce the amount of color data that needs to be transmitted or stored.

[0046] YCbCr is formed by scaling and translating YUV, where Y represents the luminance component, and Cb and Cr represent the blue and red chromaticity components, respectively.

[0047] This application takes YCbCr and RGB as examples. The differences between the image components of YCbCr and RGB are as follows: Figure 2As shown, the two can be converted to each other. Specifically, the first color space conversion refers to converting the RGB color space to the YCbCr color space. The conversion formula is as follows: ; The second color space conversion refers to converting the luminance component and the chrominance component of the second enhanced image from the YCbCr color space to the RGB color space. The conversion formula is as follows: .

[0048] Understandably, this application utilizes a pre-trained supervised model (such as LLFlow) to extract the luminance prior (Y channel) of a normally illuminated image, and adaptively fuses it with the Y channel of the enhancement result from an unsupervised model (such as Zero-DCE). This effectively compensates for defects such as uneven illumination distribution and histogram spikes while preserving the original image's natural chromaticity (Cb / Cr), making the enhanced image more closely match the detector's input expectations. Separating luminance and chromaticity not only helps reduce data volume but also ensures stable image quality performance across different application scenarios.

[0049] As an optional embodiment, fusing the luminance components of the first enhanced image and the second enhanced image to obtain a fused luminance component includes: The luminance components of the first enhanced image and the second enhanced image are weighted and fused based on a fusion network to obtain the fused luminance component. The fusion network includes a deep learning model for image fusion.

[0050] Deep learning models used for image fusion include DenseFuse, FusionGAN, DIVFusion, SD-Net, etc. This application uses SD-Net as an example for illustration.

[0051] like Figure 3 As shown, SD-Net (Squeeze-and-Decomposition Network) aims to process multimodal image fusion (such as VIF, PET-MRI) and digital image fusion (such as multi-focus, multi-exposure) in real time. It introduces compression and decomposition into image fusion, considering both the compression and decomposition processes simultaneously. Since the quality of image decomposition is related to the fused image, constraining the decomposition process can constrain the fusion process from the opposite direction. Simultaneous constraints in both directions allow the model to achieve better fusion results.

[0052] Optionally, the loss function of the fusion network includes a gradient loss function and a pixel intensity loss function. In the case of the fusion network being SD-Net, such as... Figure 3 As shown, Let be the gradient loss function. It is the pixel intensity loss function, the fused image. It is necessary to separate the input image from the input image. and Calculate the two loss functions mentioned above and set them according to a certain ratio to constrain the fused image to be more similar to which input image. and It is the re-decomposition loss, The image obtained after re-decomposition is compared with the input image for similarity calculation, followed by a reverse constraint fusion process.

[0053] Optionally, the weighted fusion of the luminance components of the first enhanced image and the second enhanced image based on the fusion network to obtain the fused luminance component includes: The first loss function weight and the second loss function weight of the fusion network are determined. The first loss function weight is used to determine the similarity between the luminance component output by the fusion network and the luminance component of the first enhanced image. The second loss function weight is used to determine the similarity between the luminance component output by the fusion network and the luminance component of the second enhanced image. The first loss function weight is greater than the second loss function weight. The fusion network performs weighted fusion of the luminance components of the first enhanced image and the second enhanced image based on the weights of the first loss function and the second loss function to obtain the fused luminance component.

[0054] When the fusion network is SD-Net, selecting appropriate loss function weights during training is crucial to ensuring the model achieves an optimal balance among different augmented image sources. 'm' represents the similarity loss weight between the fused image and the Zero-DCE augmented image, i.e., the second loss function weight, while 'n' represents the similarity loss weight between the fused image and the LLFlow augmented image, i.e., the first loss function weight. Different combinations of the first and second loss function weights directly impact the performance of the final augmented image on the object detection task. The results of the object images on the object detection task are shown in Table 1.

[0055] Table 1 Table 1 shows the average accuracy (mAP@.5) of the target image on the object detection task for different combinations of the weights of the first and second loss functions. When m is set to 0.8 and n is set to 0.2, the algorithm achieves the highest mAP@.5 value of 42.8%. In contrast, whether the weights are set equally (m=0.5, n=0.5) or favoring the LLFlow setting (m=0.2, n=0.8), the result is 42.6%, slightly lower than the former. It can be seen that although the difference in object detection results between different loss function weight settings is relatively small, the best performance is obtained when m=0.8 and n=0.2. This indicates that, under the current experimental settings, assigning higher weights to the Zero-DCE enhanced image model helps generate fused images that are more beneficial for object detection tasks. Therefore, reasonably adjusting the loss weight coefficients in the fusion module according to different application requirements and scenario characteristics is one of the key factors in improving the overall system performance.

[0056] Understandably, this application introduces SD-Net as a fusion network, and achieves high-quality fusion of illumination components through dual constraints of gradient loss and pixel intensity loss; at the same time, it designs a loss weight strategy (m=0.8, n=0.2) to prioritize the preservation of the structural details of Zero-DCE and integrate the distribution prior of LLFlow to achieve optimal detection performance.

[0057] Figure 4 This diagram illustrates a specific embodiment of the image enhancement method for object detection provided in this application. LLFlow is a supervised enhancement model, Zero-DCE is an unsupervised enhancement model, Y_fuse represents the fused luminance component, concat represents stitching, Y_1 represents the luminance component of the second enhanced image, and Y_2 represents the luminance component of the first enhanced image. Figure 3 As shown, this application uses LLFlow to enhance the image to be enhanced, performs YCbCr color space decomposition on the enhanced image, extracts the Y channel (i.e., the luminance component of the image), and fuses it with the luminance component of the image enhanced by the Zero-DCE model. The SD-Net model is then used to fuse the image. The fused Y_fuse is then combined with the Cr and Cb components of the image enhanced by the Zero-DCE model and converted back into an RGB image, thus obtaining the final enhanced image.

[0058] This application was tested against several existing low-light enhancement algorithms (SCI, EnlightenGAN, PairLIE, RUAS, LLFlow, Zero-DCE, SDAG) on different datasets (DarkFace, NOD, and custom datasets), and the target detection accuracy is shown in Table 2.

[0059] Table 2 As shown in Table 2, the proposed method achieves the best target detection accuracy on the DarkFace, NOD, and custom datasets. DarkFace and NOD are commonly used low-light target detection datasets, while the custom dataset is a low-light target detection dataset collected by the inventors. On DarkFace, the mAP@.5 reaches 42.8% (a 0.8% improvement over Zero-DCE), on NOD it reaches 52.5% (SOTA level), and on the custom dataset it reaches 30.5% (a 0.4% improvement over Zero-DCE), demonstrating its generalization and robustness. Compared to using Zero-DCE or LLFlow alone for image enhancement, this application effectively addresses the shortcomings of uneven illumination distribution, histogram artifacts, and semantic ambiguity by fusing prior brightness distribution extracted by a supervised model. This not only preserves image detail and naturalness but also effectively compensates for differences in illumination distribution, thereby significantly improving the accuracy of target detection.

[0060] Figure 5 This is a schematic diagram comparing the image enhancement effects of this application with existing algorithms, as shown below. Figure 5 As shown, this application not only performs well in object detection tasks, but also has strong generalization ability, and can stably improve the object detection effect in various scenarios. This application provides a new solution for low-light image enhancement and shows great potential in practical applications.

[0061] This application also conducts experiments on the fusion scheme of the chromaticity components of the second enhanced image, setting up three control groups: Strategy one uses all images enhanced with Zero-DCE. and The second strategy is to use LLFlow to enhance all components of the image. and For the components, strategy three uses the fusion formula given by SD-Net as follows for fusion.

[0062] Where C is the fused or , and Representing the two source images respectively or 128 is the median value of the dynamic range that is set manually.

[0063] like Figure 6As shown in Table 3, C_Zero-DCE refers to Strategy 1, C_Fuse refers to Strategy 2, and C_LLFlow refers to Strategy 3. The target detection accuracy of the three different strategies is shown in Table 3.

[0064] Table 3 from Figure 6 As can be seen from Table 3, although the LLFlow algorithm can effectively improve image brightness in some cases, the enhanced image has obvious color cast in the Cb or Cr color components. Especially when the background is dark, the color cast will introduce additional noise, which will have an adverse effect on the target detection task.

[0065] Understandably, this application considers the illumination distribution of images, analyzes the differences in illumination distribution between images enhanced by existing unsupervised enhancement algorithms and images under normal illumination, and proposes an image enhancement algorithm based on illumination component fusion. Specifically, it uses the supervised enhancement algorithm LLFlow to extract normal illumination distribution information and fuses it into the illumination components of the Zero-DCE enhanced image to compensate for the differences in illumination distribution. This makes the overall illumination distribution of the image, i.e., the pixel-level feature distribution, more consistent with images under normal illumination. Object detection detectors are generally trained on images under normal illumination; the closer the feature distribution of the target image is to the training set, the higher the object detection accuracy. Therefore, this application can theoretically improve the detection accuracy of images on object detection tasks. Extensive experimental evidence demonstrates that this application achieves the best results on object detection tasks across multiple datasets, proving its effectiveness and generalization.

[0066] As an optional embodiment, the image enhancement method for target detection provided in this application further includes: The target image is input into any preset target detector to obtain the target detection result output by the target detector.

[0067] Specifically, the target image is input into any target detector (such as YOLOv5, Faster R-CNN, etc.), and the target detector completes the target detection task without retraining the detector, thus achieving plug-and-play enhancement.

[0068] Understandably, this application utilizes deep learning to automatically learn fusion weights, without relying on manual parameters, thus avoiding color distortion and edge blurring. It achieves enhanced efficiency in low-light target detection tasks, with image quality and detection accuracy being positively correlated. It constructs a decoupled enhancement and detection architecture, outputting intermediate enhanced images without modifying or jointly training the detector. This architecture supports the integration of any detection model and possesses engineering advantages such as high modularity, reusability, and visual debugging, making it suitable for edge deployment and multi-task expansion.

[0069] The image enhancement apparatus for target detection provided in this application is described below. The image enhancement apparatus for target detection described below can be referred to in correspondence with the image enhancement method for target detection described above.

[0070] Figure 7 This is a schematic diagram of the image enhancement device for target detection provided in this application, as shown below. Figure 7 As shown, this application also provides an image enhancement apparatus for target detection, comprising: The first enhancement module 10 is used to acquire the image to be enhanced and determine a first enhanced image and a second enhanced image for supervised enhancement of the image to be enhanced; Extraction module 20 is used to extract the luminance component of the first enhanced image and extract the luminance and chrominance components of the second enhanced image; The fusion module 30 is used to fuse the luminance component of the first enhanced image and the luminance component of the second enhanced image to obtain a fused luminance component. The second enhancement module 40 is used to determine a target image for target detection based on the fused luminance component and the chrominance component of the second enhanced image.

[0071] As one embodiment, the fusion module 30 is used for: The luminance components of the first enhanced image and the second enhanced image are weighted and fused based on a fusion network to obtain the fused luminance component. The fusion network includes a deep learning model for image fusion.

[0072] As an example, the loss function of the fusion network includes a gradient loss function and a pixel intensity loss function.

[0073] As one embodiment, the fusion module 30 is used for: The first loss function weight and the second loss function weight of the fusion network are determined. The first loss function weight is used to determine the similarity between the luminance component output by the fusion network and the luminance component of the first enhanced image. The second loss function weight is used to determine the similarity between the luminance component output by the fusion network and the luminance component of the second enhanced image. The first loss function weight is greater than the second loss function weight. The fusion network performs weighted fusion of the luminance components of the first enhanced image and the second enhanced image based on the weights of the first loss function and the second loss function to obtain the fused luminance component.

[0074] As one embodiment, the first enhancement module 10 is configured to: Perform a first color space conversion on the first enhanced image and the second enhanced image; The second enhancement module 40 is used for: A second color space conversion is performed on the fused luminance component and the chrominance component of the second enhanced image to obtain the target image for target detection; The first color space conversion is related to the second color space conversion.

[0075] As one embodiment, it also includes: The target image is input into any preset target detector to obtain the target detection result output by the target detector.

[0076] It should be noted that the image enhancement device for target detection provided in this application has the same technical effects as the image enhancement method for target detection, and will not be described in detail hereafter.

[0077] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute an image enhancement method for target detection, the method including: Acquire the image to be enhanced, and determine a first enhanced image for supervised enhancement and a second enhanced image for unsupervised enhancement of the image to be enhanced; Extract the luminance component of the first enhanced image, and extract the luminance and chrominance components of the second enhanced image; The luminance components of the first enhanced image and the second enhanced image are fused to obtain a fused luminance component; Based on the fused luminance component and the chrominance component of the second enhanced image, a target image for target detection is determined.

[0078] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0079] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to perform an image enhancement method for target detection, the method comprising: Acquire the image to be enhanced, and determine a first enhanced image for supervised enhancement and a second enhanced image for unsupervised enhancement of the image to be enhanced; Extract the luminance component of the first enhanced image, and extract the luminance and chrominance components of the second enhanced image; The luminance components of the first enhanced image and the second enhanced image are fused to obtain a fused luminance component; Based on the fused luminance component and the chrominance component of the second enhanced image, a target image for target detection is determined.

[0080] In another aspect, this application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform an image enhancement method for object detection, the method comprising: Acquire the image to be enhanced, and determine a first enhanced image for supervised enhancement and a second enhanced image for unsupervised enhancement of the image to be enhanced; Extract the luminance component of the first enhanced image, and extract the luminance and chrominance components of the second enhanced image; The luminance components of the first enhanced image and the second enhanced image are fused to obtain a fused luminance component; Based on the fused luminance component and the chrominance component of the second enhanced image, a target image for target detection is determined.

[0081] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0082] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An image enhancement method for target detection, characterized in that, include: Acquire the image to be enhanced, and determine a first enhanced image for supervised enhancement and a second enhanced image for unsupervised enhancement of the image to be enhanced; Extract the luminance component of the first enhanced image, and extract the luminance and chrominance components of the second enhanced image; The luminance components of the first enhanced image and the second enhanced image are fused to obtain a fused luminance component; Based on the fused luminance component and the chrominance component of the second enhanced image, a target image for target detection is determined.

2. The image enhancement method for target detection according to claim 1, characterized in that, The step of fusing the luminance components of the first enhanced image and the second enhanced image to obtain a fused luminance component includes: The luminance components of the first enhanced image and the second enhanced image are weighted and fused based on a fusion network to obtain the fused luminance component. The fusion network includes a deep learning model for image fusion.

3. The image enhancement method for target detection according to claim 2, characterized in that, The loss function of the fusion network includes a gradient loss function and a pixel intensity loss function.

4. The image enhancement method for target detection according to claim 2, characterized in that, The weighted fusion of the luminance components of the first enhanced image and the second enhanced image based on the fusion network to obtain the fused luminance component includes: The first loss function weight and the second loss function weight of the fusion network are determined. The first loss function weight is used to determine the similarity between the luminance component output by the fusion network and the luminance component of the first enhanced image. The second loss function weight is used to determine the similarity between the luminance component output by the fusion network and the luminance component of the second enhanced image. The first loss function weight is greater than the second loss function weight. The fusion network performs weighted fusion of the luminance components of the first enhanced image and the second enhanced image based on the weights of the first loss function and the second loss function to obtain the fused luminance component.

5. The image enhancement method for target detection according to claim 1, characterized in that, Before extracting the luminance component of the first enhanced image and extracting the luminance and chrominance components of the second enhanced image, the method further includes: Perform a first color space conversion on the first enhanced image and the second enhanced image; The step of determining the target image for target detection based on the fused luminance component and the chrominance component of the second enhanced image includes: A second color space conversion is performed on the fused luminance component and the chrominance component of the second enhanced image to obtain the target image for target detection; The first color space conversion is related to the second color space conversion.

6. The image enhancement method for target detection according to any one of claims 1 to 5, characterized in that, Also includes: The target image is input into any preset target detector to obtain the target detection result output by the target detector.

7. An image enhancement device for target detection, characterized in that, include: The first enhancement module is used to acquire the image to be enhanced and determine a first enhanced image and a second enhanced image for supervised enhancement of the image to be enhanced; An extraction module is used to extract the luminance component of the first enhanced image and the luminance and chrominance components of the second enhanced image. The fusion module is used to fuse the luminance components of the first enhanced image and the second enhanced image to obtain a fused luminance component. The second enhancement module is used to determine a target image for target detection based on the fused luminance component and the chrominance component of the second enhanced image.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the image enhancement method for target detection as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the image enhancement method for target detection as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the image enhancement method for target detection as described in any one of claims 1 to 6.

Images