A non-matching low-light real image enhancement method based on multi-modal guidance

Abstract

The application discloses a kind of non-matching low-light real image enhancement methods based on multi-modal guidance, comprising the following steps: one, the acquisition of training set image;Two, construct multi-modal guidance cycle diffusion generative adversarial network;Three, the feature extraction of low-light real image and normal light image;Four, total loss function is established in the process of feature extraction in low-light real image and normal light image;Five, based on total loss function to multi-modal guidance cycle diffusion generative adversarial network training;Six, using the multi-modal guidance cycle diffusion generative adversarial network trained to single low-light image is enhanced processing.The method steps of the application are simple, and non-matching low-light real image and normal light image are trained to network by illumination perception image guidance and text perception guidance through multi-modal guidance, so that the light image output by the multi-modal guidance cycle diffusion generative adversarial network trained has good visual quality, and illumination rationality and detail texture are improved.

Description

Technical Field

[0001] This invention belongs to the field of low-light image enhancement technology, specifically relating to a method for enhancing unpaired low-light realistic images based on multimodal guidance. Background Technology

[0002] Real-world images captured in low-light environments generally suffer from insufficient illumination and loss of detail. Therefore, it is necessary to enhance low-light real-world images to recover high-quality images with uniform illumination, rich details, and natural colors. This can improve the effectiveness of low-light vision-related tasks.

[0003] Currently, most low-light realistic image enhancement methods employ supervised learning, relying on a large number of paired low-light realistic images and corresponding normal-light images (i.e., clear images) during training. However, the acquisition of large-scale, high-quality paired data faces numerous practical obstacles (such as uncontrollable lighting conditions and dynamic scene changes). Furthermore, the diversity and complexity of real-world low-light scenes limit the generalization ability of these supervised learning methods on low-light realistic images. Non-paired learning low-light realistic image enhancement methods, on the other hand, learn the true distribution characteristics of normal-light images, making the enhancement results more consistent with human visual perception, thus achieving low-light realistic image enhancement. However, these methods also have some limitations: they fail to fully explore and utilize the inherent prior information contained in the image (such as lighting distribution), their network representation capabilities are limited, and they are difficult to adapt to the complex distribution of real-world low-light scenes.

[0004] Therefore, a well-designed unpaired low-light real image enhancement method based on multimodal guidance is needed. The unpaired low-light real image and the normal light image are trained by multimodal guidance through illumination-aware image guidance and text-aware guidance. This allows the enhanced image output by the trained multimodal guidance cyclic diffusion generation adversarial network to have better visual quality, illumination rationality and detail texture improvement. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a non-paired low-light realistic image enhancement method based on multimodal guidance. The method has simple steps and reasonable design. The non-paired low-light realistic image and normal light image are trained by multimodal guidance through illumination-aware image guidance and text-aware guidance. The enhanced image output by the trained multimodal guidance cyclic diffusion generation adversarial network has good visual quality, illumination rationality and improved detail and texture.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for enhancing unpaired low-light realistic images based on multimodal guidance, the method comprising the following steps:

[0007] Step 1: Acquisition of training set images; wherein, the training set images include unpaired low-light real images and normal-light images;

[0008] Step 2: Construct a multimodal guided cyclic diffusion generative adversarial network; wherein the multimodal guided cyclic diffusion generative adversarial network includes a low-light real image enhancement and darkening branch network and a normal light image darkening and enhancement branch network, and a light-sensing image guidance module and a text-sensing guidance module are added to the low-light real image enhancement and darkening branch network and the normal light image darkening and enhancement branch network.

[0009] Step 3: Feature extraction from low-light real images and normal-light images:

[0010] Step 4: Establish a total loss function during feature extraction from low-light real images and normal-light images; wherein, the total loss function includes cycle consistency loss, identity mapping loss, adversarial loss, color constancy loss, and reflection loss. Figure 1 Consistency loss and semantic consistency loss;

[0011] Step 5: Train the multimodal guided recurrent diffusion generative adversarial network based on the total loss function to obtain the trained multimodal guided recurrent diffusion generative adversarial network;

[0012] Step 6: Use the trained multimodal guided recurrent diffusion generative adversarial network to enhance a single low-light image to obtain an enhanced image.

[0013] The above-mentioned multimodal guided unpaired low-light real image enhancement method further includes, in step one, randomly selecting real-world unpaired low-light real images and normal light images from the Unpaired Enhancement Dataset as training set images; obtaining multiple sets of training images from the training set images, each set of training images including one low-light real image and one normal light image;

[0014] The low-light real image enhancement and darkening branch network includes a first enhancement network module, a first darkening network module, and a normal light discriminator; the normal light image darkening enhancement branch network includes a second darkening network module, a second enhancement network module, and a low-light discriminator; wherein the normal light discriminator and the low-light discriminator have the same structure.

[0015] Both the first enhancement network module and the second enhancement network module include an enhancement backbone network, and both the first darkening network module and the second darkening network module include a darkening backbone network. A light-sensing image guidance module and a text-sensing guidance module are added to both the first enhancement network module and the second darkening network module. A light-sensing image guidance module is also added to both the first darkening network module and the second enhancement network module. The enhancement backbone network and the darkening backbone network have the same structure, both including a VAE encoder, a UNet, and a VAE decoder.

[0016] The illumination-aware image guidance module includes a CLIP image encoder, a linear layer, a first cross-attention layer, a second cross-attention layer, a first CLIP text encoder, and a third cross-attention layer.

[0017] The text-aware guidance module includes a BLIP model, a second CLIP text encoder, and a third CLIP text encoder.

[0018] The above-mentioned multimodal guided unpaired low-light realistic image enhancement method further includes step three, which is as follows:

[0019] Step 301: Extract the low-light realistic image After processing by the first enhancement network module, the first normal light prediction image is obtained. Among them, the first normal light prediction image The first blurred image is output after passing through the first Gaussian filter. Low-light realistic images The second blurred image is output after passing through the second Gaussian filter. ;

[0020] Normal light image After processing by the second darkening network module, the second low-light prediction image is obtained. ;

[0021] Step 302: Obtain the normal light image After processing by the first enhancement network module, the second normal light prediction image is obtained. Among them, normal light image The image is then processed by the third blurring step after passing through the first Gaussian filter. ;

[0022] Low-light realistic images After processing by the second darkening network module, the first low-light prediction image is obtained. ;

[0023] Step 303: Predict the first normal light image and low-light realistic images After processing by the first darkening network module, the third low-light prediction image is obtained. ;

[0024] Second low-light prediction image and normal light image After processing by the second enhancement network module, the third normal light prediction image is obtained. Among them, the second low-light prediction image The fourth blurred image is output after the second Gaussian filter. .

[0025] The aforementioned multimodal guided unpaired low-light real image enhancement method further includes step 301 as follows:

[0026] Step A1: Extract the low-light realistic image The first input image is converted from RGB space to YCbCr color space. The Y channel component corresponding to each pixel is extracted and subjected to max-min normalization to obtain the normalized Y channel component corresponding to each pixel. The difference between the normalized Y channel component and the Y channel component is taken as the inverted Y channel component. The inverted Y channel component corresponding to each pixel is then reverse normalized to obtain the inverted brightness map.

[0027] Step A2: Extract the low-light realistic image The corresponding inverse brightness map is processed by the illumination-sensing image guidance module in the first enhancement network module to obtain the first modal feature tensor with added illumination;

[0028] Step A3: Process the first input image through the text perception guidance module in the first enhancement network module to obtain the modal feature tensor with added text;

[0029] Step A4: The first input image, the first modal feature tensor with added illumination, and the modal feature tensor with added text are processed by the VAE encoder, UNet, and VAE decoder of the enhanced backbone network in the first enhancement network module to obtain the first output feature map, which is denoted as the first normal light prediction image. ;

[0030] Step A5: Following the methods in steps A1 to A4, extract the normal light image. The corresponding inverse brightness map is processed by the second darkening network module to obtain the second low-light prediction image. .

[0031] The aforementioned multimodal guided unpaired low-light real image enhancement method further includes step 302 as follows:

[0032] Step B1: Obtain the normal light image The second input image is converted from the RGB space to the YCbCr color space. The Y channel component corresponding to each pixel of the second input image is extracted to obtain the brightness map.

[0033] Step B2: Obtain the normal light image The corresponding brightness map is processed by the illumination-aware image guidance module in the first enhancement network module to obtain the second modal feature tensor with added illumination;

[0034] Step B3: The second input image and the modal feature tensor of the second illumination are processed by the VAE encoder, UNet and VAE decoder of the enhanced backbone network in the first enhancement network module to obtain the second output feature map, which is denoted as the second normal light prediction image. ;

[0035] Step B4: Following the methods in steps B1 to B3, extract the low-light realistic image. The corresponding brightness map is processed by the darkening backbone network in the second darkening network module and the illumination-sensing image guidance module to obtain the first low-light prediction image. .

[0036] The aforementioned multimodal guided unpaired low-light real image enhancement method further includes step 303 as follows:

[0037] Step C1: Predict the first normal light image This is recorded as the third input image, representing the low-light real image. Convert the color space from RGB to YCbCr, extract the Y channel component corresponding to each pixel of the input image, and obtain the brightness map;

[0038] Step C2: Extract the low-light realistic image The corresponding brightness map is processed by the illumination-sensing image guidance module in the first darkening network module to obtain the third modal feature tensor with added illumination;

[0039] Step C3: The third input image and the third modal feature tensor with added illumination are processed by the VAE encoder, UNet, and VAE decoder of the darkening backbone network in the first darkening network module to obtain the third output feature map, which is denoted as the third low-light prediction image. ;

[0040] Step C4: Following the methods in steps C1 to C3, process the second low-light prediction image. and normal light image The corresponding brightness map is processed by the enhanced backbone network and the illumination-sensing image guidance module in the second enhancement network module to obtain the third normal light prediction image. .

[0041] The aforementioned method for enhancing unpaired low-light realistic images based on multimodal guidance further incorporates a light-sensing image guidance module, the specific process of which is as follows:

[0042] Step A01: Record the inverted luminance map or luminance map as the luminance input image;

[0043] Step A02: The brightness input image is processed by the CLIP image encoder in the illumination-sensing image guidance module to obtain the initial feature vector;

[0044] Step A03: Input the initial feature vector into the linear layer for processing to obtain the first intermediate feature vector; reshape the first intermediate feature vector and perform tensor deformation to obtain the first intermediate feature tensor;

[0045] Step A04: Input the first intermediate feature tensor and the UNet intermediate feature map into the first cross-attention layer for processing to obtain the second intermediate feature tensor;

[0046] Step A05: Input the first intermediate feature tensor and the second intermediate feature tensor into the second cross-attention layer for processing to obtain the first feature tensor;

[0047] Step A06: If the brightness input image corresponds to a low-light real image, then the text inversion method is used to process the normal light image dataset to obtain word embedding vectors;

[0048] If the brightness input image corresponds to a normal light image, then the text inversion method is used to process the low-light real image dataset to obtain word embedding vectors;

[0049] Step A07: Process the word embedding vectors through the first CLIP text encoder to obtain the first text feature tensor; process the first text feature tensor and the UNet intermediate feature map through the third cross-attention layer to obtain the second feature tensor;

[0050] Step A08: Concatenate the first feature tensor and the second feature tensor along the sequence length dimension to obtain the modal feature tensor with illumination.

[0051] The above-mentioned multimodal guided unpaired low-light realistic image enhancement method further includes step A3, which is as follows:

[0052] Step A31: Extract the low-light realistic image The text sequence corresponding to the low-light real image is obtained by using the BLIP model;

[0053] Step A32: Replace the word "dark" with "bright" in the text sequence corresponding to the low-light real image to obtain the normal light text sequence; pass the normal light text sequence through a word segmenter to obtain the second text feature tensor;

[0054] Step A33: Process the low-light real image dataset using the text inversion method to obtain the negative cue feature vector;

[0055] Step A34: Input the second text feature tensor into the second CLIP text encoder for processing to obtain the third text feature tensor;

[0056] Step A35: Input the negative cue feature vector into the third CLIP text encoder for processing to obtain the fourth text feature tensor;

[0057] Step A36: Concatenate the third and fourth text feature tensors along the Batch dimension to obtain the fifth text feature tensor, which is denoted as the modal feature tensor of the added text.

[0058] The above-mentioned multimodal guided unpaired low-light realistic image enhancement method further includes step four, which is as follows:

[0059] Step 401, according to The circularly consistent loss is obtained. ;in, Represents a low-light realistic image and the third low-light prediction image The sum of L1 loss and LPIPS loss between them Represents normal light image and the third normal light prediction image The sum of L1 loss and LPIPS loss between them;

[0060] Step 402, according to The identity mapping loss is obtained. ;in, Represents normal light image Second normal light prediction image The sum of L1 loss and LPIPS loss between them Represents a low-light realistic image and the first low-light prediction image The sum of L1 loss and LPIPS loss between them;

[0061] Step 403, according to To obtain the countermeasure loss ;in, Represents normal light image The value output by the normal optical discriminator Represents a low-light realistic image The value output by the low-light discriminator Represents the first normal light prediction image The value output by the normal optical discriminator Indicates the second low-light prediction image The value output by the low-light discriminator;

[0062] Step 404, according to The color constant loss is obtained. ;in, Indicates the first blurred image and the third blurred image The mean square error between them This represents the fourth blurred image. With the second blurred image The mean square error between them;

[0063] Step 405, according to , to obtain reflection Figure 1 Sexual damage ;in, Indicates the second low-light prediction image The corresponding reflection pattern Represents the first normal light prediction image The corresponding reflection pattern Indicates the third low-light prediction image The corresponding reflection pattern Represents the third normal light prediction image The corresponding reflection pattern Represents a low-light realistic image The corresponding reflection pattern Represents normal light image The corresponding reflection pattern This represents the mean square error of the two reflection patterns. This represents the Manhattan distance between two reflection maps. This represents the structural similarity loss between two reflection maps;

[0064] Step 406, according to The semantic consistency loss is obtained. ;in, Represents cosine similarity. Represents a low-light realistic image The output feature map of the UNet in the first enhancement network module is obtained through the enhancement backbone network. Represents the first normal light prediction image The output feature map of the UNet in the darkening backbone network of the first darkening network module. Represents normal light image The output feature map of the UNet in the darkening backbone network of the second darkening network module. Indicates the second low-light prediction image The output feature map of UNet after passing through the enhanced backbone network in the second enhancement network module;

[0065] Step 407, according to The total loss function is obtained. .

[0066] The above-mentioned multimodal guided unpaired low-light realistic image enhancement method further includes the following step five:

[0067] The computer uses the Adam optimization algorithm, inputs multiple sets of training images, and utilizes the total loss function. The multimodal guided recurrent diffusion generative adversarial network is trained until all training images are trained to obtain a well-trained multimodal guided recurrent diffusion generative adversarial network.

[0068] Step six, the specific process is as follows:

[0069] A computer is used to input any subsequent low-light image to be processed into a trained multimodal guided recurrent diffusion generative adversarial network. After processing by the first enhancement network module in the low-light real image enhancement and darkening branch network, an enhanced output feature map is obtained, which is then the brightened image.

[0070] Compared with the prior art, the present invention has the following advantages:

[0071] 1. The method of this invention has simple steps and reasonable design. First, the training set images are acquired; second, a multimodal guided recurrent diffusion generative adversarial network is constructed; next, features are extracted from low-light real images and normal light images; then, a total loss function is established during the feature extraction process of low-light real images and normal light images; then, the multimodal guided recurrent diffusion generative adversarial network is trained; finally, the trained multimodal guided recurrent diffusion generative adversarial network is used to enhance a single low-light image and improve the image brightening effect.

[0072] 2. This invention incorporates a light-aware image guidance module and a text-aware guidance module. The light-aware image guidance module utilizes the brightness map and inverse brightness map of the original image to enhance the guidance capability of the network for lighting condition input, thereby improving the lighting rationality and detail texture of the low-light image enhancement results. The text-aware guidance module utilizes the high-level semantic information stored in the text features to constrain the network, improving the semantic consistency and visual quality of the enhancement results. Through multimodal guidance, the generated brightened image has better visual quality.

[0073] 3. The enhanced backbone network and the darkening backbone network of this invention are both VAE encoders, UNet and VAE decoders, thus making full use of the prior information contained in the diffusion model, so that the enhanced image generated by the model has better illumination rationality and visual quality, and the low-light image enhancement model has good generalization performance.

[0074] 4. The present invention sets up a first enhancement network module, a first darkening network module and a normal light discriminator, as well as a second darkening network module, a second enhancement network module and a low light discriminator, and uses cycle consistent loss to achieve effective unsupervised learning. Furthermore, a low-light real image and a normal light image are randomly selected to achieve unpaired training.

[0075] 5. This invention utilizes unpaired low-light real images and normal light images to train a multimodal guided recurrent diffusion generative adversarial network, enabling the unpaired method to process real low-light and normal light images, thereby enhancing the processing capability of low-light real images. This solves the current difficulties in collecting large-scale paired low-light real images and normal light images (i.e., clear images), as well as the problem that using synthetic paired low-light real images and clear images for brightening cannot effectively process low-light real images.

[0076] In summary, the method of this invention is simple in steps and reasonable in design. Unpaired low-light real images and normal light images are trained by multimodal guidance through illumination-aware image guidance and text-aware guidance. This allows the trained multimodal guidance to generate adversarial network output brightened images with good visual quality, illumination rationality and improved detail and texture.

[0077] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0078] Figure 1 This is a flowchart of the method of the present invention.

[0079] Figure 2 This is a schematic diagram of the structure of the multimodal guided cyclic diffusion generative adversarial network of the present invention.

[0080] Figure 3 This is a schematic diagram of the structure of the light-sensing image guidance module of the present invention.

[0081] Figure 4 This is a schematic diagram of the structure of the text perception guidance module of the present invention.

[0082] Figure 5 Images obtained by testing using the method of this invention. Detailed Implementation

[0083] like Figures 1 to 4As shown, the present invention provides a method for enhancing unpaired low-light realistic images based on multimodal guidance, comprising the following steps:

[0084] Step 1: Acquisition of training set images; wherein, the training set images include unpaired low-light real images and normal-light images;

[0085] Step 2: Construct a multimodal guided cyclic diffusion generative adversarial network; wherein the multimodal guided cyclic diffusion generative adversarial network includes a low-light real image enhancement and darkening branch network and a normal light image darkening and enhancement branch network, and a light-sensing image guidance module and a text-sensing guidance module are added to the low-light real image enhancement and darkening branch network and the normal light image darkening and enhancement branch network.

[0086] Step 3: Feature extraction from low-light real images and normal-light images:

[0087] Step 4: Establish a total loss function during feature extraction from low-light real images and normal-light images; wherein, the total loss function includes cycle consistency loss, identity mapping loss, adversarial loss, color constancy loss, and reflection loss. Figure 1 Consistency loss and semantic consistency loss;

[0088] Step 5: Train the multimodal guided recurrent diffusion generative adversarial network based on the total loss function to obtain the trained multimodal guided recurrent diffusion generative adversarial network;

[0089] Step 6: Use the trained multimodal guided recurrent diffusion generative adversarial network to enhance a single low-light image to obtain an enhanced image.

[0090] In this embodiment, in step one, unpaired real-world low-light images and normal-light images are randomly selected from the Unpaired Enhancement Dataset as training set images; multiple sets of training images are obtained from the training set images, each set of training images including one low-light image and one normal-light image;

[0091] The low-light real image enhancement and darkening branch network includes a first enhancement network module, a first darkening network module, and a normal light discriminator; the normal light image darkening enhancement branch network includes a second darkening network module, a second enhancement network module, and a low-light discriminator; wherein the normal light discriminator and the low-light discriminator have the same structure.

[0092] Both the first enhancement network module and the second enhancement network module include an enhancement backbone network, and both the first darkening network module and the second darkening network module include a darkening backbone network. A light-sensing image guidance module and a text-sensing guidance module are added to both the first enhancement network module and the second darkening network module. A light-sensing image guidance module is also added to both the first darkening network module and the second enhancement network module. The enhancement backbone network and the darkening backbone network have the same structure, both including a VAE encoder, a UNet, and a VAE decoder.

[0093] The illumination-aware image guidance module includes a CLIP image encoder, a linear layer, a first cross-attention layer, a second cross-attention layer, a first CLIP text encoder, and a third cross-attention layer.

[0094] The text-aware guidance module includes a BLIP model, a second CLIP text encoder, and a third CLIP text encoder.

[0095] In this embodiment, step three is as follows:

[0096] Step 301: Extract the low-light realistic image After processing by the first enhancement network module, the first normal light prediction image is obtained. Among them, the first normal light prediction image The first blurred image is output after passing through the first Gaussian filter. Low-light realistic images The second blurred image is output after passing through the second Gaussian filter. ;

[0097] Normal light image After processing by the second darkening network module, the second low-light prediction image is obtained. ;

[0098] Step 302: Obtain the normal light image After processing by the first enhancement network module, the second normal light prediction image is obtained. Among them, normal light image The image is then processed by the third blurring step after passing through the first Gaussian filter. ;

[0099] Low-light realistic images After processing by the second darkening network module, the first low-light prediction image is obtained. ;

[0100] Step 303: Predict the first normal light image and low-light realistic images After processing by the first darkening network module, the third low-light prediction image is obtained. ;

[0101] Second low-light prediction image and normal light image After processing by the second enhancement network module, the third normal light prediction image is obtained. Among them, the second low-light prediction image The fourth blurred image is output after the second Gaussian filter. .

[0102] In this embodiment, step 301 is as follows:

[0103] Step A1: Extract the low-light realistic image The first input image is converted from RGB space to YCbCr color space. The Y channel component corresponding to each pixel is extracted and subjected to max-min normalization to obtain the normalized Y channel component corresponding to each pixel. The difference between the normalized Y channel component and the Y channel component is taken as the inverted Y channel component. The inverted Y channel component corresponding to each pixel is then reverse normalized to obtain the inverted brightness map.

[0104] Step A2: Extract the low-light realistic image The corresponding inverse brightness map is processed by the illumination-sensing image guidance module in the first enhancement network module to obtain the first modal feature tensor with added illumination;

[0105] Step A3: Process the first input image through the text perception guidance module in the first enhancement network module to obtain the modal feature tensor with added text;

[0106] Step A4: The first input image, the first modal feature tensor with added illumination, and the modal feature tensor with added text are processed by the VAE encoder, UNet, and VAE decoder of the enhanced backbone network in the first enhancement network module to obtain the first output feature map, which is denoted as the first normal light prediction image. ;

[0107] Step A5: Following the methods in steps A1 to A4, extract the normal light image. The corresponding inverse brightness map is processed by the second darkening network module to obtain the second low-light prediction image. .

[0108] In this embodiment, step 302 is as follows:

[0109] Step B1: Obtain the normal light image The second input image is converted from the RGB space to the YCbCr color space. The Y channel component corresponding to each pixel of the second input image is extracted to obtain the brightness map.

[0110] Step B2: Obtain the normal light image The corresponding brightness map is processed by the illumination-aware image guidance module in the first enhancement network module to obtain the second modal feature tensor with added illumination;

[0111] Step B3: The second input image and the modal feature tensor of the second illumination are processed by the VAE encoder, UNet and VAE decoder of the enhanced backbone network in the first enhancement network module to obtain the second output feature map, which is denoted as the second normal light prediction image. ;

[0112] Step B4: Following the methods in steps B1 to B3, extract the low-light realistic image. The corresponding brightness map is processed by the darkening backbone network in the second darkening network module and the illumination-sensing image guidance module to obtain the first low-light prediction image. .

[0113] In this embodiment, step 303 is as follows:

[0114] Step C1: Predict the first normal light image This is recorded as the third input image, representing the low-light real image. Convert the color space from RGB to YCbCr, extract the Y channel component corresponding to each pixel of the input image, and obtain the brightness map;

[0115] Step C2: Extract the low-light realistic image The corresponding brightness map is processed by the illumination-sensing image guidance module in the first darkening network module to obtain the third modal feature tensor with added illumination;

[0116] Step C3: The third input image and the third modal feature tensor with added illumination are processed by the VAE encoder, UNet, and VAE decoder of the darkening backbone network in the first darkening network module to obtain the third output feature map, which is denoted as the third low-light prediction image. ;

[0117] Step C4: Following the methods in steps C1 to C3, process the second low-light prediction image. and normal light image The corresponding brightness map is processed by the enhanced backbone network and the illumination-sensing image guidance module in the second enhancement network module to obtain the third normal light prediction image. .

[0118] In this embodiment, the light-sensing image guidance module processes the image, and the specific process is as follows:

[0119] Step A01: Record the inverted luminance map or luminance map as the luminance input image;

[0120] Step A02: The brightness input image is processed by the CLIP image encoder in the illumination-sensing image guidance module to obtain the initial feature vector;

[0121] Step A03: Input the initial feature vector into the linear layer for processing to obtain the first intermediate feature vector; reshape the first intermediate feature vector and perform tensor deformation to obtain the first intermediate feature tensor;

[0122] Step A04: Input the first intermediate feature tensor and the UNet intermediate feature map into the first cross-attention layer for processing to obtain the second intermediate feature tensor;

[0123] Step A05: Input the first intermediate feature tensor and the second intermediate feature tensor into the second cross-attention layer for processing to obtain the first feature tensor;

[0124] Step A06: If the brightness input image corresponds to a low-light real image, then the text inversion method is used to process the normal light image dataset to obtain word embedding vectors;

[0125] If the brightness input image corresponds to a normal light image, then the text inversion method is used to process the low-light real image dataset to obtain word embedding vectors;

[0126] Step A07: Process the word embedding vectors through the first CLIP text encoder to obtain the first text feature tensor; process the first text feature tensor and the UNet intermediate feature map through the third cross-attention layer to obtain the second feature tensor;

[0127] Step A08: Concatenate the first feature tensor and the second feature tensor along the sequence length dimension to obtain the modal feature tensor with illumination.

[0128] In this embodiment, step A3 is specifically performed as follows:

[0129] Step A31: Extract the low-light realistic image The text sequence corresponding to the low-light real image is obtained by using the BLIP model;

[0130] Step A32: Replace the word "dark" with "bright" in the text sequence corresponding to the low-light real image to obtain the normal light text sequence; pass the normal light text sequence through a word segmenter to obtain the second text feature tensor;

[0131] Step A33: Process the low-light real image dataset using the text inversion method to obtain the negative cue feature vector;

[0132] Step A34: Input the second text feature tensor into the second CLIP text encoder for processing to obtain the third text feature tensor;

[0133] Step A35: Input the negative cue feature vector into the third CLIP text encoder for processing to obtain the fourth text feature tensor;

[0134] Step A36: Concatenate the third and fourth text feature tensors along the Batch dimension to obtain the fifth text feature tensor, which is denoted as the modal feature tensor of the added text.

[0135] In this embodiment, step four is as follows:

[0136] Step 401, according to The circularly consistent loss is obtained. ;in, Represents a low-light realistic image and the third low-light prediction image The sum of L1 loss and LPIPS loss between them Represents normal light image and the third normal light prediction image The sum of L1 loss and LPIPS loss between them;

[0137] Step 402, according to The identity mapping loss is obtained. ;in, Represents normal light image Second normal light prediction image The sum of L1 loss and LPIPS loss between them Represents a low-light realistic image and the first low-light prediction image The sum of L1 loss and LPIPS loss between them;

[0138] Step 403, according to To obtain the countermeasure loss ;in, Represents normal light image The value output by the normal optical discriminator Represents a low-light realistic image The value output by the low-light discriminator Represents the first normal light prediction image The value output by the normal optical discriminator Indicates the second low-light prediction image The value output by the low-light discriminator;

[0139] Step 404, according to The color constant loss is obtained. ;in, Indicates the first blurred image and the third blurred image The mean square error between them This represents the fourth blurred image. With the second blurred image The mean square error between them;

[0140] Step 405, according to , to obtain reflection Figure 1 Sexual damage ;in, Indicates the second low-light prediction image The corresponding reflection pattern Represents the first normal light prediction image The corresponding reflection pattern Indicates the third low-light prediction image The corresponding reflection pattern Indicates the third normal light prediction image The corresponding reflection pattern Represents a low-light realistic image The corresponding reflection pattern Represents normal light image The corresponding reflection pattern This represents the mean square error of the two reflection patterns. This represents the Manhattan distance between two reflection maps. This represents the structural similarity loss between two reflection maps;

[0141] Step 406, according to The semantic consistency loss is obtained. ;in, Represents cosine similarity. Represents a low-light realistic image The output feature map of the UNet in the first enhancement network module is obtained through the enhancement backbone network. Represents the first normal light prediction image The output feature map of the UNet in the darkening backbone network of the first darkening network module. Represents normal light image The output feature map of the UNet in the darkening backbone network of the second darkening network module. Indicates the second low-light prediction image The output feature map of UNet after passing through the enhanced backbone network in the second enhancement network module;

[0142] Step 407, according to The total loss function is obtained. .

[0143] In this embodiment, step five, the specific process is as follows:

[0144] The computer uses the Adam optimization algorithm, inputs multiple sets of training images, and utilizes the total loss function. The multimodal guided recurrent diffusion generative adversarial network is trained until all training images are trained to obtain a well-trained multimodal guided recurrent diffusion generative adversarial network.

[0145] Step six, the specific process is as follows:

[0146] A computer is used to input any subsequent low-light image to be processed into a trained multimodal guided recurrent diffusion generative adversarial network. After processing by the first enhancement network module in the low-light real image enhancement and darkening branch network, an enhanced output feature map is obtained, which is then the brightened image.

[0147] In this embodiment, it should be noted that, Figure 2 and Figure 3 The input shown is a low-light realistic image. .

[0148] In this embodiment, the cosine similarity acquisition in step 406 is specifically processed as follows:

[0149] Step 4061: From low-light real image Output feature map of UNet after enhancing the backbone network in the first enhancement network module The pixel values ​​of the j-th column are used to form the j-th vector. Predicted image from first normal light Output feature map of UNet after darkening backbone network in the first darkening network module The pixel values ​​of the j-th column are used to form the j-th vector. Where j is a positive integer, and the value of j ranges from 1 to J, and J is the total number of columns in the output feature map of UNet;

[0150] Step 4062: Using a computer according to... The cosine similarity between the j-th vectors is obtained. ;

[0151] Step 4063: Using a computer according to ,get and Cosine similarity between ;

[0152] Step 4064: Following the methods in steps 4061 to 4063, obtain... and Cosine similarity between .

[0153] In this embodiment, the implementation was based on the Python 3.10 programming environment and the PyTorch deep learning framework, and all experiments were performed on an NVIDIA GeForce RTX 4090 GPU. The training set images consisted of 1014 low-light real images and 1102 normal-light images. The original low-light real images and normal-light images were obtained from the training set images, and the image size was randomly cropped to 256×256, resulting in 25,000 sets of training images. During actual training, the number of training image sets can be adaptively adjusted according to requirements.

[0154] In this embodiment, during training, the low-light real image, the normal-light predicted image, the normal-light image, and the low-light predicted image are all 1×3×256×256. The brightness map and the inverted brightness map are both 1×1×256×256.

[0155] In this embodiment, it should be noted that the features of the graph are batch × number of channels × width × height, the features of the feature vector are batch × feature dimension, and the features of the feature tensor are batch × sequence length × feature dimension; the batch value is 1.

[0156] In this embodiment, the size of the initial feature vector is 1×512; the size of the first intermediate feature vector is 1×4096; the size of the first intermediate feature tensor is 1×4×1024; the size of the second intermediate feature tensor is 1×4×1024; and the size of the first feature tensor is 1×4×1024.

[0157] The word embedding vector has a size of 1×77×1024, the first text feature tensor has a size of 1×77×1024, the second feature tensor has a size of 1×77×1024, and the modal feature tensor with added illumination has a size of 1×81×1024.

[0158] The size of the second text feature tensor is 1×77×1024, and the size of the third text feature tensor is 1×77×1024;

[0159] The size of the negative prompt feature vector is 1×77×1024, the size of the fourth text feature tensor is 1×77×1024, and the size of the fifth text feature tensor is 2×77×1024.

[0160] In this embodiment, it should be noted that in step A4, the first input image is input to the VAE encoder of the enhanced backbone network in the first enhanced network module, and the first modal feature tensor with added illumination and the modal feature tensor with added text are respectively input to the first downsampled cross-attention layer in UNet;

[0161] In step B3, the second input image is input into the VAE encoder of the enhanced backbone network in the first enhancement network module, and the second modal feature tensor with added illumination is input into the first downsampled cross-attention layer in UNet;

[0162] In step C3, the third input image is input into the VAE encoder of the darkening backbone network in the first darkening network module, and the third modal feature tensor with added illumination is input into the first downsampled cross-attention layer in UNet.

[0163] In this embodiment, it should be noted that the UNet intermediate feature map refers to the low-light real image. Normal light image First normal light prediction image Or the second low-light prediction image The output after passing through the VAE encoder and the first convolutional layer of UNet.

[0164] In this embodiment, when the intermediate feature map of UNet is input into the cross-attention layer in steps A04 and A07, it needs to be flattened and processed by the linear layer to be 1×1024×1024.

[0165] In this embodiment, Figure 2 The TAG module is a text-aware guidance module, and the IAIG module is a light-aware image guidance module.

[0166] In this embodiment, it should be noted that the normal light discriminator and the low light discriminator have the same structure and are both based on the Vision-aided-gan network in the paper Ensembling Off-the-shelf Models for GAN Training, where cv_type is set to clip.

[0167] In this embodiment, the VAE encoder, UNet, and VAE decoder in the enhanced backbone network and the darkened backbone network are the diffusion model SD Turbo (v2.1).

[0168] In this embodiment, it should be noted that, Figure 2 The linear layer, the first cross-attention layer, and the second cross-attention layer in the illumination-aware image guidance module are trained. The CLIP image encoder, the first CLIP text encoder, and the third cross-attention layer in the illumination-aware image guidance module are all labeled as snowflakes and do not need to be trained.

[0169] The BLIP model, the second CLIP text encoder, and the third CLIP text encoder in the text-aware guidance module are all labeled as snowflakes and do not require training.

[0170] In this embodiment, it should be noted that, in order to efficiently utilize the pre-stored diffusion prior knowledge in SD Turbo while avoiding the high computational overhead of training from scratch, we employ a Low-Rank Adaptation (LoRA) adapter for backbone network training. UNet's LoRA rank is 128, and VAE's LoRA rank is 4. This lightweight training strategy reduces computational overhead while preserving the pre-trained model's strong representational ability of real-world scene data, ensuring the naturalness and scene adaptability of the enhanced results.

[0171] In this embodiment, it should be noted that the text inversion method can refer to the method in the paper An Image is Worth OneWord: Personalizing Text-to-Image Generation using Textual Inversion;

[0172] The CLIP image encoder, first CLIP text encoder, second CLIP text encoder, and third CLIP text encoder can all refer to the CLIP model in the paper Learning Transferable Visual Models From Natural Language Supervision.

[0173] For cross-attention layers, please refer to the paper IP-Adapter: Text Compatible Image PromptAdapter for Text-to-Image Diffusion Models.

[0174] The BLIP model can be found in the paper BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation.

[0175] In this embodiment, it should be noted that the Unpaired Enhancement Dataset can be found in the paper EnlightenGAN: Deep Light Enhancement without Paired Supervision.

[0176] In this embodiment, as Figure 5The top row is a low-light real image. Using the method of this invention, the bottom row is a brightened image. Therefore, this invention can significantly improve the enhancement effect of low-light images and also has high enhancement quality in real-world low-light images.

[0177] In this embodiment, the verification is performed on low-light image datasets, such as the LIME dataset and the NPE dataset. The LIME dataset covers various lighting scenarios, including outdoor, indoor, low light, backlight, and weak lighting. The NPE dataset consists of images collected under three low-light scenarios: cloudy, evening, and night.

[0178] In this embodiment, experiments were conducted on the NPE and LIME datasets using the method of the present invention, the UREtinex-Net method (a depth unrolling network for low-light image enhancement based on Retinex theory), the DiffUIR method (a general image restoration based on selective hourglass mapping using a diffusion model), and the UREtinex-Net++ method (an interpretable optimized unrolling network for low-light image enhancement). The quantitative comparison results are shown in Table 1.

[0179] Table 1. Quantitative comparison results of the method of the present invention with other low-light image enhancement methods.

[0180] In the quantitative comparison results of the proposed method, the UREtinex-Net method, the DiffUIR method, and the UREtinex-Net++ method on the NPE and LIME datasets, IL-NIQE, PI, and BRISQUE represent three no-reference quality evaluation indicators: integrated local natural image quality assessment, perceptual index, and no-reference image spatial quality assessment index. Lower values ​​for IL-NIQE, PI, and BRISQUE indicate better image quality. S and U represent supervised and unsupervised learning methods, respectively. It can be seen that the proposed method exhibits superior or comparable performance; therefore, the proposed method can achieve low-light image enhancement with high enhancement quality.

[0181] In this embodiment, it should be noted that the IL-NIQE evaluation metric is referenced from the paper "A feature-enriched completely blind image quality evaluator", the PI evaluation metric is referenced from the paper "Blind imagequality evaluation using perception based features", and the BRISQUE evaluation metric is referenced from the paper "No-reference image quality assessment in the spatial domain".

[0182] In summary, the method of this invention is simple in steps and reasonable in design. Unpaired low-light real images and normal light images are trained by multimodal guidance through illumination-aware image guidance and text-aware guidance. This allows the trained multimodal guidance to generate adversarial network output brightened images with good visual quality, illumination rationality and improved detail and texture.

[0183] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for enhancing unpaired low-light realistic images based on multimodal guidance, characterized in that, The method includes the following steps: Step 1: Acquisition of training set images; wherein, the training set images include unpaired low-light real images and normal-light images; Step 2: Construct a multimodal guided cyclic diffusion generative adversarial network; wherein the multimodal guided cyclic diffusion generative adversarial network includes a low-light real image enhancement and darkening branch network and a normal light image darkening and enhancement branch network, and a light-sensing image guidance module and a text-sensing guidance module are added to the low-light real image enhancement and darkening branch network and the normal light image darkening and enhancement branch network. Step 3: Feature extraction from low-light real images and normal-light images: Step 4: Establish a total loss function during feature extraction from low-light real images and normal-light images; wherein, the total loss function includes cycle consistency loss, identity mapping loss, adversarial loss, color constant loss, reflectance consistency loss and semantic consistency loss; Step 5: Train the multimodal guided recurrent diffusion generative adversarial network based on the total loss function to obtain the trained multimodal guided recurrent diffusion generative adversarial network; Step 6: Use the trained multimodal guided recurrent diffusion generative adversarial network to enhance a single low-light image to obtain an enhanced image; In step one, unpaired real-world low-light images and normal-light images are randomly selected from the Unpaired Enhancement Dataset as training set images; multiple sets of training images are obtained from the training set images, each set of training images including one low-light image and one normal-light image; The low-light real image enhancement and darkening branch network includes a first enhancement network module, a first darkening network module, and a normal light discriminator; the normal light image darkening enhancement branch network includes a second darkening network module, a second enhancement network module, and a low-light discriminator; wherein the normal light discriminator and the low-light discriminator have the same structure. Both the first enhancement network module and the second enhancement network module include an enhancement backbone network, and both the first darkening network module and the second darkening network module include a darkening backbone network. A light-sensing image guidance module and a text-sensing guidance module are added to both the first enhancement network module and the second darkening network module. A light-sensing image guidance module is also added to both the first darkening network module and the second enhancement network module. The enhancement backbone network and the darkening backbone network have the same structure, both including a VAE encoder, a UNet, and a VAE decoder. The illumination-aware image guidance module includes a CLIP image encoder, a linear layer, a first cross-attention layer, a second cross-attention layer, a first CLIP text encoder, and a third cross-attention layer. The text-aware guidance module includes a BLIP model, a second CLIP text encoder, and a third CLIP text encoder; The process is as follows, using a light-sensing image guidance module: Step A01: Record the inverted luminance map or luminance map as the luminance input image; Step A02: The brightness input image is processed by the CLIP image encoder in the illumination-sensing image guidance module to obtain the initial feature vector; Step A03: Input the initial feature vector into the linear layer for processing to obtain the first intermediate feature vector; reshape the first intermediate feature vector and perform tensor deformation to obtain the first intermediate feature tensor; Step A04: Input the first intermediate feature tensor and the UNet intermediate feature map into the first cross-attention layer for processing to obtain the second intermediate feature tensor; Step A05: Input the first intermediate feature tensor and the second intermediate feature tensor into the second cross-attention layer for processing to obtain the first feature tensor; Step A06: If the brightness input image corresponds to a low-light real image, then the text inversion method is used to process the normal light image dataset to obtain word embedding vectors; If the brightness input image corresponds to a normal light image, then the text inversion method is used to process the low-light real image dataset to obtain word embedding vectors; Step A07: Process the word embedding vectors through the first CLIP text encoder to obtain the first text feature tensor; process the first text feature tensor and the UNet intermediate feature map through the third cross-attention layer to obtain the second feature tensor; Step A08: Concatenate the first feature tensor and the second feature tensor along the sequence length dimension to obtain the modal feature tensor with illumination.

2. The method for enhancing unpaired low-light realistic images based on multimodal guidance according to claim 1, characterized in that: Step three, the specific process is as follows: Step 301: Extract the low-light realistic image After processing by the first enhancement network module, the first normal light prediction image is obtained. Among them, the first normal light prediction image The first blurred image is output after passing through the first Gaussian filter. Low-light realistic images The second blurred image is output after passing through the second Gaussian filter. ; Normal light image After processing by the second darkening network module, the second low-light prediction image is obtained. ; Step 302: Obtain the normal light image After processing by the first enhancement network module, the second normal light prediction image is obtained. Among them, normal light image The image is then processed by the third blurring step after passing through the first Gaussian filter. ; Low-light realistic images After processing by the second darkening network module, the first low-light prediction image is obtained. ; Step 303: Predict the first normal light image and low-light realistic images After processing by the first darkening network module, the third low-light prediction image is obtained. ; Second low-light prediction image and normal light image After processing by the second enhancement network module, the third normal light prediction image is obtained. Among them, the second low-light prediction image The fourth blurred image is output after the second Gaussian filter. .

3. The method for enhancing unpaired low-light realistic images based on multimodal guidance according to claim 2, characterized in that: Step 301, as follows: Step A1: Extract the low-light realistic image The first input image is converted from RGB space to YCbCr color space. The Y channel component corresponding to each pixel is extracted and subjected to max-min normalization to obtain the normalized Y channel component corresponding to each pixel. The difference between the normalized Y channel component and the Y channel component is taken as the inverted Y channel component. The inverted Y channel component corresponding to each pixel is then reverse normalized to obtain the inverted brightness map. Step A2: Extract the low-light realistic image The corresponding inverse brightness map is processed by the illumination-sensing image guidance module in the first enhancement network module to obtain the first modal feature tensor with added illumination; Step A3: Process the first input image through the text perception guidance module in the first enhancement network module to obtain the modal feature tensor with added text; Step A4: The first input image, the first modal feature tensor with added illumination, and the modal feature tensor with added text are processed by the VAE encoder, UNet, and VAE decoder of the enhanced backbone network in the first enhancement network module to obtain the first output feature map, which is denoted as the first normal light prediction image. ; Step A5: Following the methods in steps A1 to A4, extract the normal light image. The corresponding inverse brightness map is processed by the second darkening network module to obtain the second low-light prediction image. .

4. The method for enhancing unpaired low-light realistic images based on multimodal guidance according to claim 2, characterized in that: Step 302, as follows: Step B1: Obtain the normal light image The second input image is converted from the RGB space to the YCbCr color space. The Y channel component corresponding to each pixel of the second input image is extracted to obtain the brightness map. Step B2: Obtain the normal light image The corresponding brightness map is processed by the illumination-aware image guidance module in the first enhancement network module to obtain the second modal feature tensor with added illumination; Step B3: The second input image and the modal feature tensor of the second illumination are processed by the VAE encoder, UNet and VAE decoder of the enhanced backbone network in the first enhancement network module to obtain the second output feature map, which is denoted as the second normal light prediction image. ; Step B4: Following the methods in steps B1 to B3, extract the low-light realistic image. The corresponding brightness map is processed by the darkening backbone network in the second darkening network module and the illumination-sensing image guidance module to obtain the first low-light prediction image. .

5. The method for enhancing unpaired low-light realistic images based on multimodal guidance according to claim 2, characterized in that: Step 303, as follows: Step C1: Predict the first normal light image This is recorded as the third input image, representing the low-light real image. Convert the color space from RGB to YCbCr, extract the Y channel component corresponding to each pixel of the input image, and obtain the brightness map; Step C2: Extract the low-light realistic image The corresponding brightness map is processed by the illumination-sensing image guidance module in the first darkening network module to obtain the third modal feature tensor with added illumination; Step C3: The third input image and the third modal feature tensor with added illumination are processed by the VAE encoder, UNet, and VAE decoder of the darkening backbone network in the first darkening network module to obtain the third output feature map, which is denoted as the third low-light prediction image. ; Step C4: Following the methods in steps C1 to C3, process the second low-light prediction image. and normal light image The corresponding brightness map is processed by the enhanced backbone network and the illumination-sensing image guidance module in the second enhancement network module to obtain the third normal light prediction image. .

6. The method for enhancing unpaired low-light realistic images based on multimodal guidance according to claim 3, characterized in that: Step A3, the specific process is as follows: Step A31: Extract the low-light realistic image The text sequence corresponding to the low-light real image is obtained by using the BLIP model; Step A32: Replace the word "dark" with "bright" in the text sequence corresponding to the low-light real image to obtain the normal light text sequence; pass the normal light text sequence through a word segmenter to obtain the second text feature tensor; Step A33: Process the low-light real image dataset using the text inversion method to obtain the negative cue feature vector; Step A34: Input the second text feature tensor into the second CLIP text encoder for processing to obtain the third text feature tensor; Step A35: Input the negative cue feature vector into the third CLIP text encoder for processing to obtain the fourth text feature tensor; Step A36: Concatenate the third and fourth text feature tensors along the Batch dimension to obtain the fifth text feature tensor, which is denoted as the modal feature tensor of the added text.

7. The method for enhancing unpaired low-light realistic images based on multimodal guidance according to claim 2, characterized in that: Step four, the specific process is as follows: Step 401, according to The circularly consistent loss is obtained. ;in, Represents a low-light realistic image and the third low-light prediction image The sum of L1 loss and LPIPS loss between them Represents normal light image and the third normal light prediction image The sum of L1 loss and LPIPS loss between them; Step 402, according to The identity mapping loss is obtained. ;in, Represents normal light image Second normal light prediction image The sum of L1 loss and LPIPS loss between them Represents a low-light realistic image and the first low-light prediction image The sum of L1 loss and LPIPS loss between them; Step 403, according to To obtain the countermeasure loss ;in, Represents normal light image The value output by the normal optical discriminator Represents a low-light realistic image The value output by the low-light discriminator Represents the first normal light prediction image The value output by the normal optical discriminator Indicates the second low-light prediction image The value output by the low-light discriminator; Step 404, according to The color constant loss is obtained. ;in, Indicates the first blurred image and the third blurred image The mean square error between them This represents the fourth blurred image. With the second blurred image The mean square error between them; Step 405, according to The reflection map consistency loss is obtained. ;in, Indicates the second low-light prediction image The corresponding reflection pattern Represents the first normal light prediction image The corresponding reflection pattern Indicates the third low-light prediction image The corresponding reflection pattern Represents the third normal light prediction image The corresponding reflection pattern Represents a low-light realistic image The corresponding reflection pattern Represents normal light image The corresponding reflection pattern This represents the mean square error of the two reflection patterns. This represents the Manhattan distance between two reflection maps. This represents the structural similarity loss between two reflection maps; Step 406, according to The semantic consistency loss is obtained. ;in, Represents cosine similarity. Represents a low-light realistic image The output feature map of the UNet in the first enhancement network module is obtained through the enhancement backbone network. Represents the first normal light prediction image The output feature map of the UNet in the darkening backbone network of the first darkening network module. Represents normal light image The output feature map of the UNet in the darkening backbone network of the second darkening network module. Indicates the second low-light prediction image The output feature map of UNet after passing through the enhanced backbone network in the second enhancement network module; Step 407, according to The total loss function is obtained. .

8. The method for enhancing unpaired low-light realistic images based on multimodal guidance according to claim 1, characterized in that: Step 5, the specific process is as follows: The computer uses the Adam optimization algorithm, inputs multiple sets of training images, and utilizes the total loss function. The multimodal guided recurrent diffusion generative adversarial network is trained until all training images are trained to obtain a well-trained multimodal guided recurrent diffusion generative adversarial network. Step six, the specific process is as follows: A computer is used to input any subsequent low-light image to be processed into a trained multimodal guided recurrent diffusion generative adversarial network. After processing by the first enhancement network module in the low-light real image enhancement and darkening branch network, an enhanced output feature map is obtained, which is then the brightened image.

Images