Image enhancement method and device for monitoring video

Abstract

The present invention discloses an image enhancement method and device for surveillance video, which relates to the technical field of image processing. The method involves decomposing a target image into a hue component, a saturation component, and a brightness component; adaptively enhancing the saturation component to obtain a first target component; enhancing the brightness component using a pre-trained image enhancement model to obtain a second target component; and merging and transforming the hue component, the first target component, and the second target component to obtain an enhanced video frame image. By separately processing the saturation and brightness information of the image, color distortion caused by overall processing is avoided. Furthermore, by fully utilizing the deep learning model's ability to capture and enhance image details, important texture and edge details are effectively enhanced in low-light environments, making the enhanced image clearer and more delicate, thereby improving the image enhancement effect.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an image enhancement method and device for monitoring video. Background Art

[0002] Video surveillance plays a vital role in public safety, public surveillance, and home security. However, the quality of surveillance video often deteriorates significantly in low-light conditions, resulting in insufficient brightness, excessive noise, and loss of detail. This severely impacts the proper functioning of surveillance systems. To ensure the effectiveness and accuracy of surveillance video, image enhancement technology is particularly essential in low-light environments.

[0003] Traditional image enhancement methods rely on simple filters or contrast adjustments, which fail to fully capture image detail. This is especially true in low-light scenes, where important texture and detail information is easily lost. Furthermore, these methods typically process the image as a whole, failing to fully account for differences in color and brightness in low-light scenes. This can easily lead to distorted or unnatural image colors, resulting in poor image enhancement results. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem of poor enhancement effect mentioned in the above background technology, and to propose an image enhancement method and device for monitoring video.

[0005] A first aspect of the present invention provides an image enhancement method for surveillance video, the method comprising:

[0006] Obtain a target video frame image of a surveillance video and decompose it into a hue component, a saturation component, and a lightness component;

[0007] Adaptively enhancing the saturation component to obtain a first target component;

[0008] Using the brightness component as the input image of the pre-trained image enhancement model to obtain a second target component;

[0009] The hue component, the first target component, and the second target component are combined and transformed to obtain an enhanced video frame image.

[0010] Optionally, adaptively stretching the saturation component to obtain the first target component includes:

[0011] According to the maximum and minimum value normalization method, the saturation component is adaptively stretched to obtain the first target component:

[0012]

[0013] Among them, St (i, j) is the value of the first target component at position (i, j); S0(i, j) is the value of the saturation component at position (i, j); S min is the minimum value of the saturation component, S max is the maximum value of the saturation component.

[0014] Optionally, the image enhancement model includes a multi-scale feature extraction network, a feature enhancement network, a multi-scale feature fusion network, and a feature reconstruction output network; the multi-scale feature extraction network includes multiple convolutional layers, and the convolution kernels of different convolutional layers have different sizes; the feature enhancement network includes a convolutional layer for channel expansion and a convolutional layer with symmetric skip connections; wherein:

[0015] The multi-scale feature extraction network is used to receive the input image and output multiple feature maps of different sizes:

[0016]

[0017] Where Y1, Y2 and Y3 are feature maps of different sizes; x is the input image; f represents the convolution operation; K1, K2 and K3 are the convolution kernel sizes; S1, S2 and S3 are the convolution steps; σ is the activation function;

[0018] The feature enhancement network is used to perform channel expansion and feature enhancement on different feature maps to obtain multiple feature information maps F1, F2 and F3; F1, F2, F3 correspond to Y1, Y2, Y3 one by one;

[0019] The multi-scale feature fusion network is used to resize different feature information maps, perform feature fusion, and output a fused feature map;

[0020] The feature reconstruction output network is used to reconstruct the fused feature map into a single-channel map as a second target component.

[0021] Optionally, the multi-scale feature fusion network includes a downsampling layer, a first summation layer, a channel slicing layer, an attention mechanism layer, a second summation layer, an upsampling layer, and an overlay layer; the attention mechanism layer includes multiple attention blocks; wherein:

[0022] The downsampling layer is used to downsample the feature information graphs F1 and F2 respectively to obtain feature graphs G1 and G2; the sizes of G1, G2 and F3 are the same;

[0023] The first summation layer is used to add the feature map G1 and the feature map G2 to obtain a feature map G];

[0024] The channel segmentation layer is used to divide the feature map G3 into multiple feature blocks along the channel dimension;

[0025] The attention mechanism layer is used to process multiple feature blocks using multiple attention blocks to obtain multiple corresponding enhanced feature blocks; the multiple enhanced feature blocks are spliced to obtain a feature map G4;

[0026] The second summation layer is used to add the feature map G4 and the feature map F3 to obtain the feature map F4;

[0027] The upsampling layer is used to upsample the feature map F4 to obtain a feature map F5; F5 has the same size as the input image;

[0028] The superposition layer is used to splice the feature map F5 and the input image along the channel dimension to obtain a fused feature map.

[0029] Optionally, the feature reconstruction output network includes a pooling layer, a fully connected layer, a weighted fusion layer, and a single-channel reconstruction layer; wherein:

[0030] The pooling layer is used to average pool each channel of the fused feature map and output a feature vector;

[0031] The fully connected layer is configured to receive the feature vector as input and output a weight vector;

[0032] The weighted fusion layer is used to weight each channel of the fused feature map according to the weight vector to obtain a target feature map:

[0033] P t (m, n, c) = P0 (m, n, c) * A (c)

[0034] Among them, P t (m, n, c) is the value of the target feature map at position (m, n, c); P0(m, n, c) is the value of the fused feature map at position (m, n, c); A(c) is the value of the weight vector at position c;

[0035] The single-channel reconstruction layer is used to convert the target feature map into a single-channel map as the second target component.

[0036] A second aspect of the present invention provides an image enhancement device for monitoring video, the device comprising:

[0037] A conversion module is used to obtain a target video frame image of a surveillance video and decompose it into a hue component, a saturation component, and a lightness component;

[0038] a saturation adjustment module, configured to adaptively enhance the saturation component to obtain a first target component;

[0039] a luminance enhancement module, configured to use the luminance component as an input image of a pre-trained image enhancement model to obtain a second target component;

[0040] The rotation module is used to merge and transform the hue component, the first target component and the second target component to obtain an enhanced video frame image.

[0041] Optionally, the saturation adjustment module includes:

[0042] The saturation equalization calculation module is used to adaptively stretch the saturation component according to the maximum and minimum value normalization method to obtain the first target component:

[0043]

[0044] Among them, S t (i, j) is the value of the first target component at position (i, j); S0(i, j) is the value of the saturation component at position (i, j); S min is the minimum value of the saturation component, S max is the maximum value of the saturation component.

[0045] Optionally, the image enhancement model includes a multi-scale feature extraction network, a feature enhancement network, a multi-scale feature fusion network, and a feature reconstruction output network; the multi-scale feature extraction network includes multiple convolutional layers, and the convolution kernels of different convolutional layers have different sizes; the feature enhancement network includes a convolutional layer for channel expansion and a convolutional layer with symmetric skip connections; wherein:

[0046] The multi-scale feature extraction network is used to receive the input image and output multiple feature maps of different sizes:

[0047]

[0048] Where Y1, Y2 and Y3 are feature maps of different sizes; x is the input image; f represents the convolution operation; K1, K2 and K3 are the convolution kernel sizes; S1, S2 and S3 are the convolution steps; σ is the activation function;

[0049] The feature enhancement network is used to perform channel expansion and feature enhancement on different feature maps to obtain multiple feature information maps F1, F2 and F3; F1, F2, F3 correspond to Y1, Y2, Y3 one by one;

[0050] The multi-scale feature fusion network is used to resize different feature information maps, perform feature fusion, and output a fused feature map;

[0051] The feature reconstruction output network is used to reconstruct the fused feature map into a single-channel map as a second target component.

[0052] Optionally, the multi-scale feature fusion network includes a downsampling layer, a first summation layer, a channel slicing layer, an attention mechanism layer, a second summation layer, an upsampling layer, and an overlay layer; the attention mechanism layer includes multiple attention blocks; wherein:

[0053] The downsampling layer is used to downsample the feature information graphs F1 and F2 respectively to obtain feature graphs G1 and G2; the sizes of G1, G2 and F3 are the same;

[0054] The first summation layer is used to add the feature map G1 and the feature map G2 to obtain a feature map G3;

[0055] The channel segmentation layer is used to divide the feature map G3 into multiple feature blocks along the channel dimension;

[0056] The attention mechanism layer is used to process multiple feature blocks using multiple attention blocks to obtain multiple corresponding enhanced feature blocks; the multiple enhanced feature blocks are spliced to obtain a feature map G4;

[0057] The second summation layer is used to add the feature map G4 and the feature map F3 to obtain the feature map F4;

[0058] The upsampling layer is used to upsample the feature map F4 to obtain a feature map F5; F5 has the same size as the input image;

[0059] The superposition layer is used to splice the feature map F5 and the input image along the channel dimension to obtain a fused feature map.

[0060] Optionally, the feature reconstruction output network includes a pooling layer, a fully connected layer, a weighted fusion layer, and a single-channel reconstruction layer; wherein:

[0061] The pooling layer is used to average pool each channel of the fused feature map and output a feature vector;

[0062] The fully connected layer is configured to receive the feature vector as input and output a weight vector;

[0063] The weighted fusion layer is used to weight each channel of the fused feature map according to the weight vector to obtain a target feature map:

[0064] P t (m, n, c) = P0 (m, n, c) * A (c)

[0065] Among them, P t(m, n, c) is the value of the target feature map at position (m, n, c); P0(m, n, c) is the value of the fused feature map at position (m, n, c); A(c) is the value of the weight vector at position c;

[0066] The single-channel reconstruction layer is used to convert the target feature map into a single-channel map as the second target component.

[0067] Beneficial effects of the present invention:

[0068] The present invention proposes an image enhancement method for surveillance videos, which includes: obtaining a target video frame image of the surveillance video and decomposing it into a hue component, a saturation component, and a brightness component; adaptively enhancing the saturation component to obtain a first target component; using the brightness component as an input image of a pre-trained image enhancement model to obtain a second target component; and merging and converting the hue component, the first target component, and the second target component to obtain an enhanced video frame image.

[0069] By separately processing the saturation and brightness information of the image, the color distortion problem caused by overall processing is avoided; the brightness component is processed using a pre-trained image enhancement model, fully leveraging the deep learning model's ability to capture and enhance image details. In low-light environments, it can effectively enhance important texture and edge details, making the enhanced image clearer and more delicate, thereby improving the image enhancement effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The present invention will be further described below with reference to the accompanying drawings.

[0071] Figure 1 A flowchart of an image enhancement method for monitoring video is provided in an embodiment of the present invention;

[0072] Figure 2 A structural diagram of an image enhancement model is provided for an embodiment of the present invention;

[0073] Figure 3 The present invention provides a structural diagram of an image enhancement device for monitoring video. DETAILED DESCRIPTION

[0074] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0075] The embodiment of the present invention provides an image enhancement method for monitoring video. Figure 1 , Figure 1 A flowchart of an image enhancement method for surveillance video provided by an embodiment of the present invention. The method includes the following steps:

[0076] S101, obtaining a target video frame image of a surveillance video, and decomposing the target video frame image into a hue component, a saturation component, and a brightness component.

[0077] S102: Adaptively enhance the saturation component to obtain a first target component.

[0078] S103: Using the brightness component as the input image of the pre-trained image enhancement model to obtain a second target component.

[0079] S104: Merge and transform the hue component, the first target component, and the second target component to obtain an enhanced video frame image.

[0080] An image enhancement method for surveillance video provided by an embodiment of the present invention avoids color distortion caused by overall processing by separately processing the saturation and brightness information of the image. A pre-trained image enhancement model is used to process the brightness component, fully leveraging the deep learning model's ability to capture and enhance image details. Important texture and edge details can be effectively enhanced in low-light environments, making the enhanced image clearer and more delicate, thereby improving the image enhancement effect.

[0081] In one embodiment, step S102 includes:

[0082] According to the maximum and minimum normalization method, the saturation component is adaptively stretched to obtain the first target component:

[0083]

[0084] Among them, S t (i, j) is the value of the first target component at position (i, j); S0(i, j) is the value of the saturation component at position (i, j); S min is the minimum value of the saturation component, S max is the maximum value of the saturation component.

[0085] In one implementation, the saturation range is dynamically adjusted to make the colors of low-saturation areas (such as gray or blurred areas) in the image more saturated, while keeping the colors of high-saturation areas from being over-enhanced, thereby adaptively improving the image color contrast and details. Maximum and minimum value normalization can prevent the loss of color information due to low or high saturation in local areas (such as extremely dark or extremely bright areas). By stretching the entire saturation range, more areas in the image can receive appropriate color enhancement effects, improving the overall visual quality of the video.

[0086] In one embodiment, see Figure 2 , Figure 2 This is a schematic diagram of the structure of an image enhancement model provided by an embodiment of the present invention. As shown in the figure, the image enhancement model includes a multi-scale feature extraction network, a feature enhancement network, a multi-scale feature fusion network, and a feature reconstruction output network. The multi-scale feature extraction network includes multiple convolutional layers, with different convolutional kernel sizes. The feature enhancement network includes a convolutional layer for channel expansion and a convolutional layer with symmetric skip connections.

[0087] in:

[0088] The multi-scale feature extraction network receives an input image and outputs multiple feature maps of different sizes:

[0089]

[0090] Among them, Y1, Y2 and Y3 are feature maps of different sizes; x is the input image; f represents the convolution operation; K1, K2 and K3 are the convolution kernel sizes; S1, S2 and S3 are the convolution steps; σ is the activation function.

[0091] The feature enhancement network is used to perform channel expansion and feature enhancement on different feature maps respectively to obtain multiple feature information maps F1, F2 and F3; F1, F2, F3 correspond to Y1, Y2, Y3 one by one.

[0092] The multi-scale feature fusion network is used to resize different feature information maps, perform feature fusion, and output a fused feature map.

[0093] The feature reconstruction output network is used to reconstruct the fused feature map into a single channel map as the second target component.

[0094] In one implementation, Figure 2 As mentioned above, the multi-scale feature extraction network can be composed of convolutional layers co nv1 (convolution kernel size K = 7 × 7, step size S = 2), co n v 2(K=5×5, S=2) and con v3 (K = 3 × 3, S = 2). Each convolutional layer is followed by ReL u As an activation function, using convolution kernels of different sizes allows capturing image information at different levels and scales. For example, a small convolution kernel can extract local details, while a large convolution kernel can extract global structural features. This balances the preservation of details with the understanding of global information.

[0095] In one implementation, the feature enhancement network can be composed of a convolutional layer con v 4. conv 5. conv 6. conv 7. conv 8. conv 9. conv 1 0. Their convolution kernel sizes are all 3×3, and the convolution stride is all 1. co n v 4 is used for channel expansion, increasing the dimension of the feature map, allowing the network to process more image information and improving the enhancement effect. Furthermore, symmetrical skip connections connect shallow features with deep features, preserving more image details and avoiding the feature loss caused by deep convolution. This structure ensures that important edge information and detailed features are not lost during the enhancement process, improving the overall enhancement effect.

[0096] In one embodiment, a multi-scale feature fusion network includes a downsampling layer, a first summation layer, a channel slicing layer, an attention mechanism layer, a second summation layer, an upsampling layer, and an overlay layer; the attention mechanism layer includes multiple attention blocks; wherein:

[0097] The downsampling layer is used to downsample the feature information maps F1 and F2 respectively to obtain feature maps G1 and G2; the sizes of G1, G2 and F3 are the same.

[0098] The first summation layer is used to add the feature map G1 and the feature map G2 to obtain the feature map G3.

[0099] The channel segmentation layer is used to divide the feature map G3 into multiple feature blocks along the channel dimension.

[0100] The attention mechanism layer is used to process multiple feature blocks using multiple attention blocks to obtain multiple corresponding enhanced feature blocks; and the multiple enhanced feature blocks are spliced together to obtain the feature map G4.

[0101] The second summation layer is used to add the feature map G4 and the feature map F3 to obtain the feature map F4.

[0102] The upsampling layer is used to upsample the feature map F4 to obtain the feature map F5; F5 has the same size as the input image.

[0103] The overlay layer is used to concatenate the feature map F5 and the input image along the channel dimension to obtain a fused feature map.

[0104] In one implementation, the downsampling layer can be composed of convolutional layers conv11 (K=5×5, S=2), co nv12 (K=3×3, S=2), co The network consists of nv1 3 (K=3×3, S=2). Through the downsampling layer and summation structure, the network can effectively integrate feature maps from different levels and scales, enhancing the context perception ability of the model.

[0105] In one implementation, the attention mechanism layer uses multiple attention blocks to process feature blocks, which can automatically focus on important feature areas and can use transforme r Attention blocks in the network. This selective attention mechanism enables the network to dynamically identify and enhance details or features that are important to the image, thereby improving the overall enhancement effect, especially in complex or low-contrast images.

[0106] In one implementation, upsampling may use a bilinear interpolation method.

[0107] In one embodiment, the feature reconstruction output network includes a pooling layer, a fully connected layer, a weighted fusion layer, and a single-channel reconstruction layer; wherein:

[0108] The pooling layer is used to average pool each channel of the fused feature map and output a feature vector.

[0109] The fully connected layer receives the feature vector as input and outputs the weight vector.

[0110] The weighted fusion layer is used to weight each channel of the fusion feature map according to the weight vector to obtain the target feature map:

[0111] P t (m, n, c) = P0 (m, n, c) * A (c)

[0112] Among them, P t (m, n, c) is the value of the target feature map at position (m, n, c); P0(m, n, c) is the value of the fused feature map at position (m, n, c); A(c) is the value of the weight vector at position c.

[0113] The single-channel reconstruction layer is used to convert the target feature map into a single-channel map as the second target component.

[0114] In one implementation, global average pooling is used to reduce the dimensionality of the feature map and aggregate the spatial information of the feature channels. A fully connected layer estimates the weight coefficients of different channels, assigning greater weights to important features. After this fully connected layer, sigmoid is used as the activation function. The weight coefficients are multiplied by the original feature channels to obtain the target feature map. This is then passed through a convolutional layer (conv14) (K=3×3, S=1) to further enhance the network's feature representation capabilities. This is then processed through a convolutional layer (conv15) (K=1×1, S=1) to obtain a single-channel map.

[0115] The embodiment of the present invention provides an image enhancement device for monitoring video. Figure 3 , Figure 3 This is a structural diagram of an image enhancement device for monitoring video provided by an embodiment of the present invention. The device includes:

[0116] The conversion module is used to obtain the target video frame image of the surveillance video and decompose it into a hue component, a saturation component and a brightness component.

[0117] The saturation adjustment module is used to adaptively enhance the saturation component to obtain a first target component.

[0118] The luminance enhancement module is used to use the luminance component as the input image of the pre-trained image enhancement model to obtain a second target component.

[0119] The rotation module is used to merge and transform the hue component, the first target component and the second target component to obtain an enhanced video frame image.

[0120] An image enhancement device for surveillance video provided by an embodiment of the present invention avoids color distortion caused by overall processing by separately processing the saturation and brightness information of the image. A pre-trained image enhancement model is used to process the brightness component, fully leveraging the deep learning model's ability to capture and enhance image details. Important texture and edge details can be effectively enhanced in low-light environments, making the enhanced image clearer and more delicate, thereby improving the image enhancement effect.

[0121] In one embodiment, the saturation adjustment module includes:

[0122] The saturation equalization calculation module is used to adaptively stretch the saturation component according to the maximum and minimum value normalization method to obtain the first target component:

[0123]

[0124] Among them, S t(i, j) is the value of the first target component at position (i, j); S0(i, j) is the value of the saturation component at position (i, j); S min is the minimum value of the saturation component, S max is the maximum value of the saturation component.

[0125] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for enhancing an image of a surveillance video, characterized in that: The method comprises: Obtain a target video frame image of a surveillance video and decompose it into a hue component, a saturation component, and a lightness component; Adaptively enhancing the saturation component to obtain a first target component; Using the brightness component as an input image of a pre-trained image enhancement model to obtain a second target component; Merging and transforming the hue component, the first target component, and the second target component to obtain an enhanced video frame image; The image enhancement model includes a multi-scale feature extraction network, a feature enhancement network, a multi-scale feature fusion network and a feature reconstruction output network; the multi-scale feature extraction network includes multiple convolutional layers, and the convolution kernels of different convolutional layers have different sizes; the feature enhancement network includes a convolutional layer for channel expansion and a convolutional layer with symmetrical skip connections; in: The multi-scale feature extraction network is used to receive the input image and output multiple feature maps of different sizes: Where Y1, Y2 and Y3 are feature maps of different sizes; x is the input image; f represents the convolution operation; K1, K2 and K3 are the convolution kernel sizes; S1, S2 and S3 are the convolution steps; σ is the activation function; The feature enhancement network is used to perform channel expansion and feature enhancement on different feature maps respectively to obtain multiple feature information maps F1, F2 and F3; F1, F2, F3 correspond to Y1, Y2, Y3 one by one; the feature enhancement network includes seven convolutional layers, namely conv4, conv5, conv6, conv7, conv8, conv9 and conv10; the convolution kernel size of each convolutional layer is 3×3, and the step size is 1; among them, conv4 receives the output of the upper network as input; conv4, conv5, conv6, conv7, conv8, conv9 and conv10 are connected in sequence, and symmetric connections are added between conv5 and conv10, conv6 and conv9, and conv7 and conv8 respectively; The multi-scale feature fusion network is used to resize different feature information maps, perform feature fusion, and output a fused feature map; The feature reconstruction output network is used to reconstruct the fused feature map into a single-channel map as a second target component.

2. The image enhancement method for monitoring video according to claim 1, characterized in that: Adaptively stretching the saturation component to obtain the first target component includes: According to the maximum and minimum value normalization method, the saturation component is adaptively stretched to obtain the first target component: Among them, S t (i, j) is the value of the first target component at position (i, j); S0(i, j) is the value of the saturation component at position (i, j); S min is the minimum value of the saturation component, S max is the maximum value of the saturation component.

3. The image enhancement method for monitoring video according to claim 1, characterized in that: The multi-scale feature fusion network includes a downsampling layer, a first summation layer, a channel segmentation layer, an attention mechanism layer, a second summation layer, an upsampling layer, and an overlay layer; the attention mechanism layer includes multiple attention blocks; wherein: The downsampling layer is used to downsample the feature information graphs F1 and F2 respectively to obtain feature graphs G1 and G2; the sizes of G1, G2 and F3 are the same; The first summation layer is used to add the feature map G1 and the feature map G2 to obtain a feature map G3; The channel segmentation layer is used to divide the feature map G3 into multiple feature blocks along the channel dimension; The attention mechanism layer is used to process multiple feature blocks using multiple attention blocks to obtain multiple corresponding enhanced feature blocks; the multiple enhanced feature blocks are spliced to obtain a feature map G4; The second summation layer is used to add the feature map G4 and the feature map F3 to obtain the feature map F4; The upsampling layer is used to upsample the feature map F4 to obtain a feature map F5; F5 has the same size as the input image; The superposition layer is used to splice the feature map F5 and the input image along the channel dimension to obtain a fused feature map.

4. The image enhancement method for monitoring video according to claim 1, characterized in that: The feature reconstruction output network includes a pooling layer, a fully connected layer, a weighted fusion layer and a single-channel reconstruction layer; wherein: The pooling layer is used to average pool each channel of the fused feature map and output a feature vector; The fully connected layer is configured to receive the feature vector as input and output a weight vector; The weighted fusion layer is used to weight each channel of the fused feature map according to the weight vector to obtain a target feature map: P t (m,n,c)=P0(m,n,c)*A(c) Among them, P t (m, n, c) is the value of the target feature map at position (m, n, c); P0(m, n, c) is the value of the fused feature map at position (m, n, c); A(c) is the value of the weight vector at position c; The single-channel reconstruction layer is used to convert the target feature map into a single-channel map as the second target component.

5. An image enhancement device for monitoring video, characterized in that: The device comprises: A conversion module is used to obtain a target video frame image of a surveillance video and decompose it into a hue component, a saturation component, and a lightness component; a saturation adjustment module, configured to adaptively enhance the saturation component to obtain a first target component; a luminance enhancement module, configured to use the luminance component as an input image of a pre-trained image enhancement model to obtain a second target component; a rotation module, configured to combine and transform the hue component, the first target component, and the second target component to obtain an enhanced video frame image; The image enhancement model includes a multi-scale feature extraction network, a feature enhancement network, a multi-scale feature fusion network and a feature reconstruction output network; the multi-scale feature extraction network includes multiple convolutional layers, and the convolution kernels of different convolutional layers have different sizes; the feature enhancement network includes a convolutional layer for channel expansion and a convolutional layer with symmetrical skip connections; in: The multi-scale feature extraction network is used to receive the input image and output multiple feature maps of different sizes: Where Y1, Y2 and Y3 are feature maps of different sizes; x is the input image; f represents the convolution operation; K1, K2 and K3 are the convolution kernel sizes; S1, S2 and S3 are the convolution steps; σ is the activation function; The feature enhancement network is used to perform channel expansion and feature enhancement on different feature maps respectively to obtain multiple feature information maps F1, F2 and F3; F1, F2, F3 correspond to Y1, Y2, Y3 one by one; the feature enhancement network includes seven convolutional layers, namely conv4, conv5, conv6, conv7, conv8, conv9 and conv10; the convolution kernel size of each convolutional layer is 3×3, and the step size is 1; among them, conv4 receives the output of the upper network as input; conv4, conv5, conv6, conv7, conv8, conv9 and conv10 are connected in sequence, and symmetric connections are added between conv5 and conv10, conv6 and conv9, and conv7 and conv8 respectively; The multi-scale feature fusion network is used to resize different feature information maps, perform feature fusion, and output a fused feature map; The feature reconstruction output network is used to reconstruct the fused feature map into a single-channel map as a second target component.

6. The image enhancement device for monitoring video according to claim 5, characterized in that: The saturation adjustment module includes: The saturation equalization calculation module is used to adaptively stretch the saturation component according to the maximum and minimum value normalization method to obtain the first target component: Among them, S t (i, j) is the value of the first target component at position (i, j); S0(i, j) is the value of the saturation component at position (i, j); S min is the minimum value of the saturation component, S max is the maximum value of the saturation component.

7. The image enhancement device for monitoring video according to claim 5, characterized in that: The multi-scale feature fusion network includes a downsampling layer, a first summation layer, a channel segmentation layer, an attention mechanism layer, a second summation layer, an upsampling layer, and an overlay layer; the attention mechanism layer includes multiple attention blocks; wherein: The downsampling layer is used to downsample the feature information graphs F1 and F2 respectively to obtain feature graphs G1 and G2; the sizes of G1, G2 and F3 are the same; The first summation layer is used to add the feature map G1 and the feature map G2 to obtain a feature map G3; The channel segmentation layer is used to divide the feature map G3 into multiple feature blocks along the channel dimension; The attention mechanism layer is used to process multiple feature blocks using multiple attention blocks to obtain multiple corresponding enhanced feature blocks; the multiple enhanced feature blocks are spliced to obtain a feature map G4; The second summation layer is used to add the feature map G4 and the feature map F3 to obtain the feature map F4; The upsampling layer is used to upsample the feature map F4 to obtain a feature map F5; F5 has the same size as the input image; The superposition layer is used to splice the feature map F5 and the input image along the channel dimension to obtain a fused feature map.

8. The image enhancement device for monitoring video according to claim 5, characterized in that: The feature reconstruction output network includes a pooling layer, a fully connected layer, a weighted fusion layer and a single-channel reconstruction layer; wherein: The pooling layer is used to average pool each channel of the fused feature map and output a feature vector; The fully connected layer is configured to receive the feature vector as input and output a weight vector; The weighted fusion layer is used to weight each channel of the fused feature map according to the weight vector to obtain a target feature map: P t (m,n,c)=P0(m,n,c)*A(c) Among them, P t (m, n, c) is the value of the target feature map at position (m, n, c); P0(m, n, c) is the value of the fused feature map at position (m, n, c); A(c) is the value of the weight vector at position c; The single-channel reconstruction layer is used to convert the target feature map into a single-channel map as the second target component.

Images