A zero-shot demand low-light image enhancement method and system

Abstract

The application provides a low-illumination image enhancement method and system without sample requirements, and relates to the technical field of image processing. The low-illumination image to be enhanced is separated into a hue channel image, a saturation channel image and a brightness channel image; the interpolated brightness channel image is input into a convolutional neural network to obtain correction parameters; the correction parameters include shadow amount, highlight amount and correction amount; the original brightness channel image is corrected by using the correction parameters to obtain a corrected brightness channel image; the corrected brightness channel image is spliced with the original hue channel image and the original saturation channel image, and is converted to obtain a reconstructed enhanced image; wherein the convolutional neural network is trained under unsupervised conditions by minimizing neighborhood consistency loss and exposure control loss. The application does not depend on any prior image examples or prior training, and realizes low-illumination image enhancement with good effects in zero sample and multiple types of scenes.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a low-light image enhancement method and system with zero sample requirements. Background Technology

[0002] Low-light image enhancement refers to the technical process of improving the visual quality of images taken in low-light conditions, which suffer from low brightness, blurred details, and significant noise, making them clearer, brighter, and more in line with human visual perception. Low-light image enhancement technology has an urgent application need in many scenarios such as security monitoring and night photography.

[0003] To address the challenges of low-light images, various image enhancement methods have emerged in the current technology. Traditional methods, such as histogram equalization, can improve brightness to some extent, but they are prone to causing local overexposure or loss of detail. Supervised methods based on deep learning rely on a large number of pairs of low-light and normal-light images for training. However, in practical applications, obtaining such labeled data is costly and difficult, limiting their widespread application. While some unsupervised methods have eliminated the dependence on paired data, they still fall short in terms of the precision of the enhancement effect, especially in balancing the preservation of details in shadow areas and the suppression of overexposure in highlight areas. They cannot simultaneously meet the requirements of overall image brightness improvement, rich detail, and color fidelity. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a low-light image enhancement method and system with zero-sample requirements. It achieves effective low-light image enhancement in various scenarios without relying on any prior image examples or prior training.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a low-light image enhancement method with zero sample requirement, comprising: The low-light image to be enhanced is separated into hue channel map, saturation channel map and luminance channel map; The brightness channel map is interpolated and input into a convolutional neural network to obtain correction parameters; the correction parameters include shadow amount, highlight amount, and correction amount. A brightness correction curve is constructed using the correction parameters, and the original brightness channel map is adaptively corrected to obtain the corrected brightness channel map. The corrected luminance channel image is then stitched together with and converted from the original hue and saturation channel images to obtain the reconstructed enhanced image. The convolutional neural network is trained under unsupervised conditions by minimizing neighborhood consistency loss and exposure control loss. The neighborhood consistency loss is used to maintain the consistency of brightness differences in local regions before and after enhancement, and the exposure control loss is used to control the brightness of each local region to approach the standard gray level.

[0006] Preferably, the step of separating the low-light image to be enhanced into a hue channel image, a saturation channel image, and a luminance channel image specifically includes: acquiring the low-light image to be enhanced, converting it from the RGB color gamut space to the HSV color gamut space, and obtaining the hue channel image, saturation channel image, and luminance channel image respectively.

[0007] Preferably, the brightness correction curve is as follows: ; in, and These are the input and output brightness values; increment function. Defined as , , , These are preset parameters, with values ​​ranging from [0, 0.5]. This is the shadow amount, used to shift underexposed areas to a well-exposed level; High light intensity is used to suppress brightness overflow in overexposed areas; This is a correction factor used to stretch and compress the entire image, enhancing details in low-grayscale areas.

[0008] Preferably, in the unsupervised training of the convolutional neural network, parameter self-learning is achieved by calculating neighborhood consistency loss and exposure control loss based on the original luminance channel map and the corrected luminance channel map.

[0009] Preferably, the neighborhood consistency loss is specifically: ; in, This refers to the number of local regions, where a single local region is set to a non-overlapping pixel area of ​​size 4x4. Based on the region It consists of eight adjacent areas: top, bottom, left, right, top left, top right, bottom left, and bottom right. and These represent the average brightness values ​​of local regions in the enhanced image and the input image, respectively. These represent the average brightness values ​​of the local region centered on region i and the local region j adjacent to region i in the enhanced image, respectively. This represents the average brightness value of the local region centered on region i and the local region j adjacent to region i in the input image.

[0010] Preferably, the exposure control loss specifically refers to: ; in, This indicates the number of non-overlapping local regions of size 8×8. and These represent the average brightness values ​​of local regions in the enhanced image and the input image, respectively. Used to shrink Deviation from medium gray; Indicates the activation function; This represents the average brightness value of the k-th local region in the input image; This represents the average brightness value of the k-th local region in the enhanced image.

[0011] Preferably, the conversion is from the HSV color gamut space to the RGB color gamut space.

[0012] Secondly, the present invention provides a low-light image enhancement system with zero sample requirement, comprising: The image preprocessing module is used to separate the low-light image to be enhanced into a hue channel map, a saturation channel map, and a brightness channel map; The correction parameter acquisition module is used to interpolate the brightness channel map, input it into the convolutional neural network, and obtain correction parameters. The correction parameters include shadow amount, highlight amount, and correction amount. The convolutional neural network is trained under unsupervised conditions by minimizing neighborhood consistency loss and exposure control loss. The neighborhood consistency loss is used to maintain the consistency of brightness difference in local areas of the image before and after enhancement, and the exposure control loss is used to control the brightness of each local area to approach the gray level in the standard. The correction module is used to construct a brightness correction curve using the correction parameters, and to adaptively correct the original brightness channel map to obtain the corrected brightness channel map. The reconstruction module is used to stitch and convert the corrected luminance channel image with the original hue channel image and saturation channel image to obtain the reconstructed enhanced image.

[0013] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the zero-sample-requirement low-light image enhancement method described in the first aspect.

[0014] Fourthly, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the zero-sample-requirement low-light image enhancement method described in the first aspect.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention separates the image into hue, saturation, and brightness channels, and adaptively corrects the brightness channel by constructing a brightness correction curve using the shadow amount, highlight amount, and correction amount output by the convolutional neural network. This effectively improves the brightness of low-light areas and suppresses brightness overflow in overexposed areas, while accurately preserving the original color and saturation information of the image, achieving synergistic optimization of brightness and color, and enhancing image details and visual effects. Secondly, the convolutional neural network is trained under unsupervised conditions by minimizing neighborhood consistency loss and exposure control loss. It does not rely on reference images, reducing dependence on labeled data and saving data acquisition and labeling costs. At the same time, neighborhood consistency loss maintains the consistency of brightness differences in local areas before and after enhancement, ensuring the coherence of the image neighborhood, and exposure control loss makes the brightness of local areas approach the gray level in the standard, effectively avoiding underexposure or overexposure problems.

[0016] (2) This invention introduces the HSV color space, separates the light brightness information and color information of RGB images, and only analyzes and processes the brightness. Under the premise of ensuring the correlation of color information, it greatly reduces the complexity of the model, improves the inference speed, realizes the effective mapping of images to brightness fitting curves under various lighting conditions, and has good environmental adaptability.

[0017] (3) The present invention proposes a deep learning network for target images only, which realizes effective mapping of images to brightness correction curves under various lighting conditions, and restores the image to the imaging effect under normal lighting conditions through the curve.

[0018] (4) The present invention designs a loss evaluation function based on neighborhood consistency to evaluate the brightness enhancement result in the absence of a reference image. By iteratively minimizing this function, the parameters of the brightness correction curve are optimized.

[0019] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0021] Figure 1 A main flowchart of a low-light image enhancement method with zero sample requirement provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a low-light image enhancement method with zero sample requirement provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the brightness correction curve provided in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Example 1 like Figure 1 As shown, this embodiment discloses a low-light image enhancement method with zero sample requirement, including the following steps: S1: Separate the low-light image to be enhanced into a hue channel map, a saturation channel map, and a brightness channel map; S2: Interpolate the brightness channel map and input it into a convolutional neural network to obtain correction parameters; the correction parameters include shadow amount, highlight amount and correction amount; S3: Construct a brightness correction curve using the correction parameters, and perform adaptive correction on the original brightness channel map to obtain the corrected brightness channel map; S4: The corrected luminance channel image is spliced ​​and converted with the original hue channel image and saturation channel image to obtain the reconstructed enhanced image.

[0024] Next, combined Figure 2 This embodiment provides a detailed description of a low-light image enhancement method with zero sample requirements.

[0025] First, acquire the low-light image to be enhanced, and segment its RGB image by separating the light intensity information and color information.

[0026] Specifically, an RGB image is converted into a hue channel map using a general conversion formula from RGB to HSV (Hue, Saturation, Value) color gamut space. Saturation channel diagram and brightness channel image I V .

[0027] In the HSV color gamut space, the V channel represents the brightness of light. The brightness value of any pixel in the image in the HSV color gamut space is calculated using formula (1). v The value range is [0,1], where r, g, and b represent the original RGB channel values ​​of the pixel.

[0028] (1) like Figure 2 As shown, the brightness channel image of the image is obtained. I V Because the luminance V depends on the R, G, and B values ​​equally, traditional image processing functions such as convolution and equalization can be implemented by processing the luminance value.

[0029] Design a brightness correction curve as follows: Figure 3 As shown, the horizontal axis represents the original brightness value, and the vertical axis represents the optimized adjustment value. The brightness correction curve can be determined by the amount of shadow. High light intensity and correction amount The three parameters are quantized and expressed as: (2) in and These are the input and output brightness values ​​in the image. Increment function. Defined as , where the default parameter ( This ensures that the value of falls within the range [0, 0.5]. Shading amount It can help move underexposed areas to a well-exposed level, while high light intensity Overexposed areas can be helped by suppressing brightness overflow in overexposed areas. This stretches and compresses the entire image, further enhancing the details in low-grayscale areas of the target image.

[0030] This embodiment proposes three parameters: shadow amount, highlight amount, and gamma correction amount. Shadow amount improves underexposed areas, highlight amount corrects overexposed areas, and gamma correction amount stretches and compresses the overall image brightness. It can specifically optimize uneven illumination, achieving adaptive brightness adjustment without relying on a reference image. In particular, it enhances detail features in low grayscale areas, making key information such as foreign objects and textures in dark areas clearer, providing a high-quality brightness foundation for subsequent object detection and improving the image's recognizability under complex lighting conditions.

[0031] Furthermore, an effective mapping from image to brightness correction curve is performed under various lighting conditions, and this curve is used to restore the image to the imaging effect under normal lighting conditions: Design a Convolutional Neural Network (CNN) to process an input image. The optimal luminance correction curve is estimated in the luminance channel, and then the luminance is adjusted using this curve. The lighting effect is restored, and its structure is as follows: Figure 2 As shown.

[0032] First, the luminance channel image is interpolated. I V The specifications were adjusted to obtain a brightness image with a fixed size of 256x256. Then, through calculations using 7 convolutional layers and 5 fully connected layers, three intermediate estimates of shadow amount, highlight amount, and correction amount are output, respectively. / / This indicates that the intermediate estimate is substituted into the brightness correction curve. After correction, the following results were obtained. .

[0033] Furthermore, a loss evaluation function is used to evaluate the brightness enhancement results without a reference image, thereby optimizing the parameters of the fitted curve: A loss function is designed to evaluate the exposure recovery effect, and its input is the normalized brightness map before and after correction. and The output is an evaluation value that includes both neighborhood consistency loss and exposure control loss. By iteratively minimizing these values, the image can be obtained. Corresponding optimal parameters , and .

[0034] Domain consistency loss This encourages neighborhood coherence of the enhanced image by preserving the differences between neighboring regions of the input image and its enhanced version. (3) in, This refers to the number of local regions, with each local region set to a non-overlapping pixel area of ​​size 4x4. Based on the region The center consists of eight adjacent areas (top, bottom, left, right, top left, top right, bottom left, bottom right). and These represent the average brightness values ​​of local regions in the enhanced version and the input image, respectively. This represents the average brightness value of the local region centered on region i in the enhanced image; This represents the average local brightness value of region j adjacent to region i in the enhanced image. This represents the average brightness value of a local region centered at region i in the input image. This represents the average local brightness value of region j adjacent to region i in the input image.

[0035] In this embodiment, the neighborhood consistency loss encourages the enhanced image to maintain neighborhood coherence by calculating the brightness difference between adjacent regions of the enhanced image and the original image. It compares the average brightness difference between the central region and eight adjacent regions in 4x4 non-overlapping pixel units, avoiding image distortion caused by sudden changes in local brightness. This loss effectively maintains the overall structural integrity of the image, preventing problems such as edge blurring and detail breakage during enhancement, ensuring that low-light images retain natural spatial coherence after brightness enhancement, and improving visual effects and data reliability.

[0036] Exposure control loss Referring to assumptions widely used in the field of exposure fusion, the exposure rationality of each region is evaluated by the difference between the average brightness value and the mid-gray (0.5) of the local area, thereby suppressing underexposed / overexposed areas.

[0037] (4) in, This indicates the number of non-overlapping local regions of size 8×8. and Similarly, these represent the average brightness values ​​of local areas in the enhanced version and the input image, respectively. encourage Get as close to mid-gray as possible so that both underexposed and overexposed areas can be adjusted to a well-exposed level. Meanwhile, Also encourage inherit The relationship between and 0.5, if Greater than (less than) 0.5 It is also often larger (smaller) than 0.5, thus ensuring Darker / brighter areas in The center is still relatively dark / bright; This represents the average brightness value of the k-th local region in the input image; This represents the average brightness value of the k-th local region in the enhanced image.

[0038] In this embodiment, the exposure control loss assesses the reasonableness of exposure by using the average brightness difference of the 8x8 non-overlapping regions. This suppresses underexposed or overexposed areas while preserving the original image's brightness and darkness. It avoids the problem of an overall image being too bright or too dark after brightness adjustment, while ensuring a natural imaging logic where dark areas remain dark and bright areas remain bright, restoring images in unevenly lit scenes to normal exposure levels. This loss accurately controls the overall exposure, providing image data with balanced brightness and rich detail for target detection.

[0039] The total loss can be expressed as: (5) in = 0.65、 = 0.35, which are the weights of the loss.

[0040] In this embodiment, the total loss is achieved by fusing neighborhood consistency loss and exposure control loss through weighted subtraction. This approach highlights the core requirement of exposure optimization while also considering the integrity of image structure. The subtraction logic balances the constraints of the two losses, preventing biases caused by over-optimization of a single loss. For example, it prevents over-correction of exposure from disrupting neighborhood consistency, or excessive structural constraints from affecting the exposure improvement effect. This achieves the dual goals of brightness optimization and structure preservation, enabling the model to iterate to the optimal brightness correction curve and effectively improve the accuracy and stability of low-light image enhancement.

[0041] In each training iteration, the enhancement effect of the brightness channel is evaluated using a loss function, and the enhancement value is estimated by minimizing the loss. , , The parameterized optimal luminance correction curve is used to adjust the original luminance map. Correction is performed to obtain image, Hue channel map corresponding to the original image Saturation channel diagram Together, they are converted to RGB space using a universal HSV to RGB conversion formula to obtain the final reconstructed image under normal lighting conditions. .

[0042] This specific embodiment optimizes brightness by separating image channels, constructs a correction curve using a convolutional neural network to output shadow amount, highlight amount, and correction amount, and achieves unsupervised parameter optimization by combining neighborhood consistency loss and exposure control loss. Compared with traditional low-light enhancement methods, this invention achieves good image reconstruction results in various scene types through deep learning without relying on any prior image examples or prior training.

[0043] Meanwhile, this specific embodiment autonomously estimates the optimal brightness correction curve parameters from the brightness channel of the image to be enhanced using a CNN model. Combined with the total loss function comprised of neighborhood consistency loss and exposure control loss, it can iteratively optimize the correction parameters even without a reference image. This avoids the dependence on reference samples used in traditional methods, reducing data acquisition and annotation costs. It also adapts to diverse and complex scenarios such as uneven outdoor lighting, ensuring that images with varying low-light conditions can be accurately restored to normal lighting effects. This provides high-quality image data for target detection under zero-sample conditions, improving the practicality and scene adaptability of the technology.

[0044] Example 2 This embodiment provides a low-light image enhancement system with zero sample requirement, including: The image preprocessing module is used to separate the low-light image to be enhanced into a hue channel map, a saturation channel map, and a brightness channel map; The correction parameter acquisition module is used to interpolate the brightness channel map, input it into the convolutional neural network, and obtain correction parameters. The correction parameters include shadow amount, highlight amount, and correction amount. The convolutional neural network is trained under unsupervised conditions by minimizing neighborhood consistency loss and exposure control loss. The neighborhood consistency loss is used to maintain the consistency of brightness difference in local areas of the image before and after enhancement, and the exposure control loss is used to control the brightness of each local area to approach the gray level in the standard. The correction module is used to construct a brightness correction curve using the correction parameters, and to adaptively correct the original brightness channel map to obtain the corrected brightness channel map. The reconstruction module is used to stitch and convert the corrected luminance channel image with the original hue channel image and saturation channel image to obtain the reconstructed enhanced image.

[0045] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a zero-sample-requirement low-light image enhancement method as described in Embodiment 1 above.

[0046] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the zero-sample-requirement low-light image enhancement method described in Embodiment 1 above.

[0047] The steps or modules involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A low-light image enhancement method with zero sample requirement, characterized in that, include: The low-light image to be enhanced is separated into hue channel map, saturation channel map and luminance channel map; The brightness channel map is interpolated and input into a convolutional neural network to obtain correction parameters; the correction parameters include shadow amount, highlight amount, and correction amount. A luminance correction curve is constructed using the aforementioned correction parameters, and the original luminance channel map is adaptively corrected to obtain the corrected luminance channel map; the luminance correction curve is specifically as follows: ； in, and These are the input and output brightness values; increment function. Defined as , , , These are preset parameters, with values ​​ranging from [0, 0.5]. This is the shadow amount, used to shift underexposed areas to a well-exposed level; High light intensity is used to suppress brightness overflow in overexposed areas; The correction amount is used to stretch and compress the entire image, enhancing the details in low grayscale areas of the image; The corrected luminance channel image is then stitched together with and converted from the original hue and saturation channel images to obtain the reconstructed enhanced image. The convolutional neural network is trained under unsupervised conditions by minimizing neighborhood consistency loss and exposure control loss. The neighborhood consistency loss is used to maintain the consistency of brightness differences in local regions before and after enhancement, and the exposure control loss is used to control the brightness of each local region to approach the gray level in the standard. The neighborhood consistency loss is specifically as follows: ； in, This refers to the number of local regions, where a single local region is set to a non-overlapping pixel area of ​​size 4x4. Based on the region It consists of eight adjacent areas: top, bottom, left, right, top left, top right, bottom left, and bottom right. and These represent the average brightness values ​​of local regions in the enhanced image and the input image, respectively. These represent the average brightness values ​​of the local region centered on region i and the local region j adjacent to region i in the enhanced image, respectively. This represents the average brightness value of a local region centered on region i and a local region j adjacent to region i in the input image. The exposure control loss specifically refers to: ； in, This indicates the number of non-overlapping local regions of size 8×8. and These represent the average brightness values ​​of local regions in the enhanced image and the input image, respectively. Used to shrink Deviation from medium gray; Indicates the activation function; This represents the average brightness value of the k-th local region in the input image; This represents the average brightness value of the k-th local region in the enhanced image.

2. The low-light image enhancement method with zero sample requirement as described in claim 1, characterized in that, The step of separating the low-light image to be enhanced into a hue channel image, a saturation channel image, and a luminance channel image specifically includes: acquiring the low-light image to be enhanced, converting it from the RGB color gamut space to the HSV color gamut space, and obtaining the hue channel image, saturation channel image, and luminance channel image respectively.

3. The low-light image enhancement method with zero sample requirement as described in claim 1, characterized in that, In the unsupervised training of the convolutional neural network, the neighborhood consistency loss and exposure control loss are calculated based on the original luminance channel map and the corrected luminance channel map to achieve parameter self-learning.

4. The low-light image enhancement method with zero sample requirement as described in claim 1, characterized in that, The conversion is from the HSV color gamut space to the RGB color gamut space.

5. A low-light image enhancement system with zero sample requirement, characterized in that, include: The image preprocessing module is used to separate the low-light image to be enhanced into a hue channel map, a saturation channel map, and a brightness channel map; The correction parameter acquisition module is used to interpolate the brightness channel map, input it into a convolutional neural network, and obtain correction parameters. The correction parameters include shadow amount, highlight amount, and correction amount. The convolutional neural network is trained under unsupervised conditions by minimizing neighborhood consistency loss and exposure control loss. The neighborhood consistency loss is used to maintain the consistency of brightness differences in local regions before and after enhancement, and the exposure control loss is used to control the brightness of each local region to approach the standard gray level. Specifically, the neighborhood consistency loss is: ； in, This refers to the number of local regions, where a single local region is set to a non-overlapping pixel area of ​​size 4x4. Based on the region It consists of eight adjacent areas: top, bottom, left, right, top left, top right, bottom left, and bottom right. and These represent the average brightness values ​​of local regions in the enhanced image and the input image, respectively. These represent the average brightness values ​​of the local region centered on region i and the local region j adjacent to region i in the enhanced image, respectively. This represents the average brightness value of a local region centered on region i and a local region j adjacent to region i in the input image. The exposure control loss specifically refers to: ； in, This indicates the number of non-overlapping local regions of size 8×8. and These represent the average brightness values ​​of local regions in the enhanced image and the input image, respectively. Used to shrink Deviation from medium gray; Indicates the activation function; This represents the average brightness value of the k-th local region in the input image; This represents the average brightness value of the k-th local region in the enhanced image; The correction module is used to construct a brightness correction curve using the correction parameters, and to adaptively correct the original brightness channel map to obtain the corrected brightness channel map; the brightness correction curve is specifically as follows: ； in, and These are the input and output brightness values; increment function. Defined as , , , These are preset parameters, with values ​​ranging from [0, 0.5]. This is the shadow amount, used to shift underexposed areas to a well-exposed level; High light intensity is used to suppress brightness overflow in overexposed areas; The correction amount is used to stretch and compress the entire image, enhancing the details in low grayscale areas of the image; The reconstruction module is used to stitch and convert the corrected luminance channel image with the original hue channel image and saturation channel image to obtain the reconstructed enhanced image.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in a low-light image enhancement method with zero sample requirements as described in any one of claims 1-4.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the low-light image enhancement method with zero sample requirement as described in any one of claims 1-4.

Images