A Scene Text Super-Resolution Method and System Based on Text Prior and Stationary Wavelet Domain Transform

Abstract

This invention discloses a method and system for scene text super-resolution based on text prior and stationary wavelet domain transform. The method includes: constructing an initial super-resolution image generation neural network model and inputting historical low-resolution text images into the model to obtain historical high-resolution text images; applying stationary wavelet transform to the historical high-resolution text images to obtain low-frequency sub-bands and several high-frequency sub-bands, and performing loss calculation to construct a frequency domain loss function; constructing a discriminator network model and calculating an adversarial loss function; constructing a generator total loss function based on the frequency domain loss and adversarial loss, and using the total loss to update and iterate the initial super-resolution image generation neural network model and the discriminator network model to obtain a super-resolution image generation neural network model; acquiring the low-resolution text image to be processed, and using the super-resolution image generation neural network model to process the low-resolution text image to be processed to obtain the super-resolution result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision and image processing, and particularly relates to a scene text super-resolution method and system based on text priori and stationary wavelet domain transformation. BACKGROUND

[0002] With the development of deep learning, scene text recognition is gradually applied in automatic driving, smart city and artificial intelligence, etc. The clarity and readability of scene text directly affect the effect of text recognition, especially in the case of low resolution, super-resolution technology is particularly important. Scene text super-resolution faces multiple challenges:

[0003] (1) High frequency information loss: low resolution text images often lack the necessary high frequency details, resulting in blurred generated images, making it difficult to accurately identify characters and information; (2) Complex background interference: text images are usually in complex backgrounds, and traditional super-resolution methods are difficult to effectively separate text and background when processing, affecting the readability of the image; (3) Blurring and artifact problem: existing deep learning models may introduce blurring and artifacts when performing super-resolution reconstruction, further reducing the readability of the text; (4) Diversity and deformation: scene text has various fonts, sizes and writing styles, and existing super-resolution algorithms often struggle to cope with these diversities, resulting in ineffective recovery of text features.

[0004] The core goal of text image super-resolution is to improve the clarity of low resolution text images, making them have higher readability in scene text recognition, document image processing and other tasks. However, existing methods still face problems such as high frequency detail loss, artifact interference and unstable text region structure when dealing with real degraded text images, affecting the readability of the text and the accuracy of subsequent recognition tasks. SUMMARY

[0005] The present application aims to solve the problems of the prior art and provides the following solutions:

[0006] A scene text super-resolution method based on text priori and stationary wavelet domain transformation, comprising the following steps:

[0007] An initial super-resolution image generation neural network model is constructed, and historical low resolution text images are input into the initial super-resolution image generation neural network model to obtain historical high resolution text images;

[0008] The historical high resolution text images are applied to stationary wavelet transform to obtain low frequency subbands and a plurality of high frequency subbands, and a frequency domain loss function is constructed based on loss calculation of the historical high resolution text images, the low frequency subbands, the high frequency subbands and real high resolution text images;

[0009] The discriminator network model is constructed, the high-frequency subband and the corresponding binary text mask are spliced as input, and an adversarial loss function is calculated;

[0010] The generator total loss function is constructed based on the frequency domain loss and the adversarial loss, and the initial super-resolution image generation neural network model and the discriminator network model are updated and iterated by using the total loss, to obtain a super-resolution image generation neural network model;

[0011] A low-resolution text image to be processed is obtained, and the super-resolution image generation neural network model is used to process the low-resolution text image to be processed, to obtain a super-resolution result.

[0012] Preferably, the initial super-resolution image generation neural network model comprises a text prior generation unit, an image feature extraction unit, a text prior interpretation unit and an image reconstruction unit;

[0013] The text prior generation unit is used to extract character recognition probability information from an original low-resolution image using a pre-trained CRNN network, to generate a text semantic probability map;

[0014] The image feature extraction unit is used to extract features of the original low-resolution image using a multi-layer convolutional neural network, to obtain a spatial feature map;

[0015] The text prior interpretation unit is used to splice the text semantic probability map and the spatial feature map in the channel dimension, and perform convolution fusion operation, to obtain a fused feature map after fusion;

[0016] The image reconstruction unit is composed of a plurality of up-sampling units and residual blocks, and a high-resolution image is obtained by processing the fused feature map.

[0017] Preferably, the method for constructing the frequency domain loss function comprises:

[0018] The generated historical high-resolution text image I SR is decomposed into four frequency band subgraphs by stationary wavelet transform:

[0019] I SWT = SWT(I SR ) = {I LL , I LH , I HL , I HH}

[0020] Wherein, I LL represents a low-frequency subband, I LH , I HL and I HH represent high-frequency subbands;

[0021] The historical high-resolution text image is compared with the corresponding part of the real high-resolution text image on each sub-band, and loss calculation is performed to construct a frequency domain loss function:

[0022]

[0023] wherein λ L1 represents a low-frequency loss weighting coefficient, λ L2 represents a high-frequency loss weighting coefficient, represents a low-frequency sub-band of the high-definition image, represents a low-frequency sub-band of the super-resolution image, represents a high-frequency sub-band of the high-definition image, represents a high-frequency sub-band of the super-resolution image.

[0024] Preferably, the discriminator network model comprises a spatial attention unit, a feature extraction unit and a classification unit.

[0025] The spatial attention unit is composed of two parallel branches of maximum pooling operation and average pooling operation, for respectively performing pooling processing and splicing enhancement on the input feature map in the spatial dimension to obtain an enhanced spatial attention feature map.

[0026] The feature extraction unit is composed of 4 sequentially stacked deep residual blocks, for encoding the enhanced spatial attention feature map to obtain an encoded intermediate feature map.

[0027] The classification unit is composed of a global average pooling layer and 2 fully connected layers, for compressing the encoded intermediate feature map into a one-dimensional vector, and linearly mapping the one-dimensional vector to obtain a real value without activation function processing, and performing adversarial loss calculation based on the real value.

[0028] Preferably, the method for calculating the adversarial loss function comprises:

[0029]

[0030] wherein L adv represents the adversarial loss represents the expectation under the joint distribution of the generated image and its binary mask, Softplus represents the Softplus activation function, D represents the discriminator, F SR represents the high-frequency sub-band of the super-resolution image, M SR represents the binary mask of the super-resolution image.

[0031] Preferably, the method for constructing the total loss function of the generator comprises:

[0032] Based on structural similarity loss, pixel-level L1 reconstruction loss, and text-aware loss, the basic loss function of the generator is constructed as follows:

[0033] L SR =αL TSSIM +βL1+γL TP

[0034] Among them, L SR L represents the basic loss function. TSSIM L1 represents the structural similarity loss, and L2 represents the pixel-level L1 reconstruction loss. TP Let α represent the weighting coefficient of the structural similarity loss, β represent the weighting coefficient of the pixel-level L1 reconstruction loss, and γ represent the weighting coefficient of the text perception loss.

[0035] Based on the frequency domain loss function, the basic loss function, and the adversarial loss function, the total loss function is constructed as follows:

[0036] L total =L SWT +L SR +ηL adv

[0037] Among them, L total Let represent the total loss function, and η represent the adversarial loss weight.

[0038] The present invention also provides a scene text super-resolution system based on text prior and stationary wavelet domain transform. The system applies the method described in any of the above-mentioned methods and includes: a model building module, a first loss function building module, a second loss function building module, a model update module, and a super-resolution processing module.

[0039] The model building module is used to build an initial super-resolution image generation neural network model, and input historical low-resolution text images into the initial super-resolution image generation neural network model to obtain historical high-resolution text images;

[0040] The first loss function construction module is used to apply stationary wavelet transform to the historical high-resolution text image to obtain low-frequency sub-bands and several high-frequency sub-bands, and to perform loss calculation based on the historical high-resolution text image, the low-frequency sub-bands, the high-frequency sub-bands and the real high-resolution text image to construct a frequency domain loss function;

[0041] The second loss function construction module is used to construct a discriminator network model, taking the high-frequency sub-band and the corresponding binary text mask as input, and calculating the adversarial loss function;

[0042] The model update module constructs a generator total loss function based on the frequency domain loss and the adversarial loss, and uses the total loss to update and iterate the initial super-resolution image generation neural network model and the discriminator network model to obtain the super-resolution image generation neural network model.

[0043] The super-resolution processing module is used to acquire the low-resolution text image to be processed, and to process the low-resolution text image using the super-resolution image generation neural network model to obtain the super-resolution result.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] This invention effectively enhances the model's ability to model the structure of text regions by introducing prior information of text generated based on a recognition network; it decomposes the image into multiple frequency sub-bands using a stationary wavelet transform method, achieving separate modeling of structural and edge information and improving the quality of detail restoration; the constructed wavelet domain discriminator network model combines text masking and high-frequency sub-band features, and introduces a channel attention mechanism to improve the ability to discriminate text structural regions; and by optimizing multiple loss functions, it effectively suppresses artifacts and significantly improves the quality of the generated image and the recognizability of the text. Attached Figure Description

[0046] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram of the discriminator network model in an embodiment of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Example 1

[0052] In this embodiment, as Figure 1 As shown, a scene text super-resolution method based on text prior and stationary wavelet domain transform includes the following steps:

[0053] S1. Construct an initial super-resolution image generation neural network model, and input historical low-resolution text images into the initial super-resolution image generation neural network model to obtain historical high-resolution text images.

[0054] The initial super-resolution image generation neural network model includes: a text prior generation unit, an image feature extraction unit, a text prior interpretation unit, and an image reconstruction unit. The text prior generation unit uses a pre-trained CRNN network to extract character recognition probability information from the original low-resolution image, generating a text semantic probability map. The image feature extraction unit uses a multi-layer convolutional neural network to extract features from the original low-resolution image, obtaining a spatial feature map. The text prior interpretation unit concatenates the text semantic probability map and the spatial feature map along the channel dimension and performs a convolutional fusion operation to obtain a fused feature map. The image reconstruction unit consists of several upsampling units and residual blocks; by processing the fused feature map, a high-resolution image is obtained.

[0055] In this embodiment, (1) Text prior generation module: uses a pre-trained convolutional recurrent neural network (CRNN) as the feature extraction backbone, receives the input low-resolution text image, outputs character-level recognition probability information, and generates a text semantic probability map with dimensions corresponding to the image size, which is used to provide text prior information for subsequent module fusion. (2) Image feature extraction module: consists of multiple convolutional neural networks, the input is the original low-resolution image, and the output is a multi-channel spatial feature map, which is used to characterize low-level visual features such as edges and textures of the image. The module may include standard convolutional layers, batch normalization layers and nonlinear activation functions. (3) Text prior interpretation module: receives the semantic probability map output from the text prior generation module and the spatial feature map output from the image feature extraction module, concatenates them in the channel dimension and inputs them into several convolutional layers, outputting a fused joint feature map, which jointly represents the spatial and semantic features of the image. (4) Image reconstruction module: consists of multiple upsampling units and residual blocks, and the input is the fused feature map output by the text prior interpretation module. The upsampling structure includes transposed convolutional layers or pixel rearrangement operation modules, which are used to enlarge the spatial dimension of the feature map to twice that of the original image; residual blocks are used for deep feature modeling, and the output is the final generated high-resolution image.

[0056] A large number of low-resolution and high-resolution image pairs are used as training data. These data are used to generate high-resolution outputs through a feedforward neural network. Subsequently, various losses between the predicted and real high-resolution images are calculated, including pixel-level L1 loss and structural similarity loss (SSIM). The network adjusts its internal parameters through the backpropagation algorithm to reduce the difference between the predicted and real images. During the optimization process, the Adam optimizer is used to adjust the learning rate and other hyperparameters to obtain the best training results.

[0057] The system generates a high-resolution text image with twice the resolution of the original image from a low-resolution RGB text image through neural network processing. First, the system adjusts the received low-resolution RGB text image and its binary mask to the standard size required for network processing. The input image first passes through a feature extraction module, which uses a convolutional neural network to extract image features. Simultaneously, a text prior generation module uses a pre-trained CRNN network to extract character recognition probabilities from the image as text prior information and fuses them with the image features as input to the backbone network. The image features processed by feature extraction and text prior fusion are then fed into an image reconstruction module. This module utilizes pixel shuffle upsampling technology to enhance image resolution. Furthermore, the module includes residual connections to help recover more high-frequency details. Finally, this module outputs a high-resolution text image with twice the width and height of the original input image.

[0058] S2. Apply stationary wavelet transform to the historical high-resolution text image to obtain a low-frequency sub-band and several high-frequency sub-bands. Calculate the loss based on the historical high-resolution text image, the low-frequency sub-band, the high-frequency sub-band, and the real high-resolution text image, and construct a frequency domain loss function.

[0059] In this embodiment, the method for constructing the frequency domain loss function includes: taking the generated historical high-resolution text image I... SR The stationary wavelet transform decomposes the image into four frequency band sub-images:

[0060] I SWT =SWT(I SR )={I LL I LH I HL I HH}

[0061] Among them, I LL Indicates low-frequency subband, I LH I HL and I HH Representing high-frequency sub-bands; comparing the corresponding portions of historical high-resolution text images and real high-resolution text images in each sub-band, and calculating the loss to construct a frequency domain loss function:

[0062]

[0063] Where, λ L1 λ represents the low-frequency loss weighting coefficient. L2 Represents the high-frequency loss weighting coefficient. The low-frequency subband representing high-definition images. This represents the low-frequency subband of the super-resolution image. High-frequency subband representing high-definition images, This represents the high-frequency subband of a super-resolution image.

[0064] S3. Construct a discriminator network model, using the high-frequency sub-band and the corresponding binary text mask as input, and calculate the adversarial loss function.

[0065] The four sub-bands are concatenated along the channel dimension to form a high-frequency feature set. The original low-resolution image is converted into a grayscale image, and its grayscale mean is calculated and used as a binarization threshold to generate a binary mask for the text region. Finally, the high-frequency feature set and the text mask are further concatenated along the channel dimension to form the final discriminator input.

[0066] Discriminator network model such as Figure 2 As shown, the system includes a spatial attention unit, a feature extraction unit, and a classification unit. The spatial attention unit consists of two parallel branches: max pooling and average pooling. These branches pool the input feature map in the spatial dimension, resulting in two single-channel feature maps. These feature maps are then concatenated in the channel dimension and input into a single convolutional layer, outputting a spatial attention weight map. Finally, this weight map is multiplied element-wise with the input tensor to generate an enhanced spatial attention feature map, with the output size matching the input. The feature extraction unit consists of four sequentially stacked depth residual blocks, denoted as DB1, DB2, DB3, and DB4. Each residual block contains two convolutional layers and a cross-layer connection path. The input is the feature map output from the previous module, and the output is a high-dimensional convolutional feature map. The channel dimension can gradually increase between blocks, and the spatial size can be adjusted according to the convolution settings. This module ultimately outputs a set of encoded intermediate feature maps. The classification unit consists of a global average pooling layer (GAP) and two fully connected layers (FC1 and FC2). The GAP layer is used to compress the encoded intermediate feature map into a one-dimensional vector in the spatial dimension. This vector is then input into FC1 and FC2 for linear mapping, and finally outputs a real value that has not been processed by the activation function. This value is used for adversarial loss calculation, and the output dimension is a single scalar.

[0067] Methods for calculating adversarial loss functions include:

[0068]

[0069] Among them, L advIndicating resistance to loss, Let F represent the expectation of the joint distribution of the generated image and its binary mask, Softplus represent the Softplus activation function, D represent the discriminator, and F represent the expected value. SR M represents the high-frequency subband of a super-resolution image. SR A binary mask representing a super-resolution image.

[0070] S4. Construct a generator total loss function based on frequency domain loss and adversarial loss, and use the total loss to update and iterate the initial super-resolution image generation neural network model and discriminator network model to obtain the super-resolution image generation neural network model.

[0071] Methods for constructing the generator's total loss function include:

[0072] Based on structural similarity loss, pixel-level L1 reconstruction loss, and text-aware loss, the basic loss function of the generator is constructed as follows:

[0073] L SR =αL TSSIM +βL1+γL TP Among them, L SR L represents the basic loss function. TSSIM L1 represents the structural similarity loss, and L2 represents the pixel-level L1 reconstruction loss. TP Let α represent the weighting coefficient of the structural similarity loss, β represent the weighting coefficient of the pixel-level L1 reconstruction loss, and γ represent the weighting coefficient of the text perception loss. Based on the frequency domain loss function, the basic loss function, and the adversarial loss function, a total loss function is constructed:

[0074] L total =L SWT +L SR +ηL adv

[0075] Among them, L total Let represent the total loss function, and η represent the adversarial loss weight, which is used to control the strength of the generator's response to the discriminator's feedback.

[0076] In each training iteration, the discriminator is first fixed and the generator parameters are updated. Then the generator is fixed and the discriminator parameters are updated. The above processes obtain the loss value through forward propagation and calculate the gradient and complete the parameter update through the backpropagation algorithm. The optimizer is Adam, and the learning rate and update frequency are adjusted according to experience. Throughout the training process, the generator and discriminator are updated in turn.

[0077] Then, feedforward-backward propagation is performed to train the initial super-resolution image generation neural network model: a low-resolution text image is used as the input to the model, which is then passed to the generator network and the discriminator network in turn. The forward inference process is performed to generate a super-resolution image with twice the resolution of the original image and output the image's authenticity judgment result. Based on the differences between the generated image and the real image in the pixel domain, structural domain, text semantic domain, and frequency domain, combined with the discriminator's discrimination output, the joint loss function of the generator and the adversarial loss function of the discriminator are calculated. The loss terms include generation loss, stationary wavelet transform subband loss, and adversarial loss. Through the backpropagation algorithm, the loss gradients of the generator and discriminator are passed from the output layer to the input layer layer by layer, and the parameters of each network are updated in turn, completing one round of feedforward-backward propagation training process.

[0078] Determine whether the current training state meets the termination condition. If the generator's total loss function on the validation set tends to converge in multiple consecutive training rounds, or the number of training rounds reaches the preset maximum number of rounds, then the model training is considered complete, the training process is terminated, and the final model parameters are saved. If the condition is not met, proceed to step 2, iteratively update the network parameters until the training termination criterion is met, and the final super-resolution image generation neural network model is obtained.

[0079] S5. Obtain the low-resolution text image to be processed, and use the super-resolution image generation neural network model to process the low-resolution text image to obtain the super-resolution result.

[0080] Example 2

[0081] In this embodiment, a scene text super-resolution system based on text prior and stationary wavelet domain transform includes: a model building module, a first loss function building module, a second loss function building module, a model update module, and a super-resolution processing module.

[0082] The model building module is used to build an initial super-resolution image generation neural network model and input historical low-resolution text images into the initial super-resolution image generation neural network model to obtain historical high-resolution text images.

[0083] The first loss function construction module applies stationary wavelet transform to historical high-resolution text images to obtain low-frequency sub-bands and several high-frequency sub-bands. Based on the historical high-resolution text images, low-frequency sub-bands, high-frequency sub-bands, and real high-resolution text images, loss calculation is performed to construct a frequency domain loss function.

[0084] The second loss function construction module is used to build the discriminator network model. It takes the high-frequency subband and the corresponding binary text mask as input and calculates the adversarial loss function.

[0085] The model update module constructs a generator total loss function based on frequency domain loss and adversarial loss, and uses the total loss to update and iterate the initial super-resolution image generation neural network model and discriminator network model to obtain the super-resolution image generation neural network model.

[0086] The super-resolution processing module is used to acquire the low-resolution text image to be processed, and to use the super-resolution image generation neural network model to process the low-resolution text image to obtain the super-resolution result.

[0087] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A scene text super-resolution method based on text prior and stationary wavelet domain transform, characterized in that, Includes the following steps: An initial super-resolution image generation neural network model is constructed, and historical low-resolution text images are input into the initial super-resolution image generation neural network model to obtain historical high-resolution text images; The historical high-resolution text image is subjected to stationary wavelet transform to obtain a low-frequency sub-band and several high-frequency sub-bands. Based on the historical high-resolution text image, the low-frequency sub-band, the high-frequency sub-band and the real high-resolution text image, loss calculation is performed to construct a frequency domain loss function. Construct a discriminator network model, using the high-frequency sub-band and the corresponding binary text mask as input, and calculate the adversarial loss function; Based on the frequency domain loss and the adversarial loss, a generator total loss function is constructed. The total loss is then used to update and iterate the initial super-resolution image generation neural network model and the discriminator network model to obtain the super-resolution image generation neural network model. A low-resolution text image to be processed is obtained, and the low-resolution text image to be processed is processed using the super-resolution image generation neural network model to obtain a super-resolution result; The initial super-resolution image generation neural network model includes: a text prior generation unit, an image feature extraction unit, a text prior interpretation unit, and an image reconstruction unit; The text prior generation unit is used to extract character recognition probability information from the original low-resolution image using a pre-trained CRNN network to generate a text semantic probability map. The image feature extraction unit is used to extract features from the original low-resolution image using a multi-layer convolutional neural network to obtain a spatial feature map; The text prior interpretation unit is used to concatenate the text semantic probability map and the spatial feature map in the channel dimension and perform a convolutional fusion operation to obtain a fused feature map. The image reconstruction unit consists of several upsampling units and residual blocks. By processing the fused feature map, a high-resolution image is obtained. The method for constructing the frequency domain loss function includes: The generated historical high-resolution text image I SR The stationary wavelet transform decomposes the image into four frequency band sub-images: Among them, I LL Indicates low-frequency subband, I LH I HL and I HH Indicates high-frequency sub-band; By comparing the corresponding portions of the historical high-resolution text image with the real high-resolution text image in each sub-band and calculating the loss, a frequency domain loss function is constructed: Where, λ L1 λ represents the low-frequency loss weighting coefficient. H1 Represents the high-frequency loss weighting coefficient. The low-frequency subband representing high-definition images. This represents the low-frequency subband of the super-resolution image. High-frequency subband representing high-definition images, This represents a specific high-frequency subband of a super-resolution image.

2. The scene text super-resolution method based on text prior and stationary wavelet domain transform according to claim 1, characterized in that, The discriminator network model includes: a spatial attention unit, a feature extraction unit, and a classification unit; The spatial attention unit consists of two parallel branches: max pooling and average pooling, which are used to perform pooling processing on the input feature map in the spatial dimension and then concatenate and enhance it to obtain the enhanced spatial attention feature map. The feature extraction unit consists of four sequentially stacked depth residual blocks, which are used to encode the enhanced spatial attention feature map to obtain the encoded intermediate feature map. The classification unit consists of a global average pooling layer and two fully connected layers, used to compress the encoded intermediate feature map into a one-dimensional vector, and then perform a linear mapping on the one-dimensional vector to obtain a real value without activation function processing, and to calculate the adversarial loss based on the real value.

3. The scene text super-resolution method based on text prior and stationary wavelet domain transform according to claim 2, characterized in that, The method for calculating the adversarial loss function includes: Among them, L adv Indicating resistance to loss, Let F represent the expectation of the joint distribution of the generated image and its binary mask, Softplus represent the Softplus activation function, D represent the discriminator, and F represent the expected value. SR M represents the set of all high-frequency subbands of a super-resolution image. SR A binary mask representing a super-resolution image.

4. The scene text super-resolution method based on text prior and stationary wavelet domain transform according to claim 3, characterized in that, The method for constructing the total loss function of the generator includes: Based on structural similarity loss, pixel-level L1 reconstruction loss, and text-aware loss, the basic loss function of the generator is constructed as follows: Among them, L SR L represents the basic loss function. TSSIM L1 represents the structural similarity loss, and L2 represents the pixel-level L1 reconstruction loss. TP Let α represent the weighting coefficient of the structural similarity loss, β represent the weighting coefficient of the pixel-level L1 reconstruction loss, and γ represent the weighting coefficient of the text perception loss. Based on the frequency domain loss function, the basic loss function, and the adversarial loss function, the total loss function is constructed as follows: Among them, L total Let represent the total loss function, and η represent the adversarial loss weight.

5. A scene text super-resolution system based on text prior and stationary wavelet domain transform, wherein the system applies the method described in any one of claims 1-4, characterized in that, include: The module includes a model building module, a first loss function building module, a second loss function building module, a model update module, and a super-resolution processing module. The model building module is used to build an initial super-resolution image generation neural network model, and input historical low-resolution text images into the initial super-resolution image generation neural network model to obtain historical high-resolution text images; The first loss function construction module is used to apply stationary wavelet transform to the historical high-resolution text image to obtain low-frequency sub-bands and several high-frequency sub-bands, and to perform loss calculation based on the historical high-resolution text image, the low-frequency sub-bands, the high-frequency sub-bands and the real high-resolution text image to construct a frequency domain loss function; The second loss function construction module is used to construct a discriminator network model, taking the high-frequency sub-band and the corresponding binary text mask as input, and calculating the adversarial loss function; The model update module constructs a generator total loss function based on the frequency domain loss and the adversarial loss, and uses the total loss to update and iterate the initial super-resolution image generation neural network model and the discriminator network model to obtain the super-resolution image generation neural network model. The super-resolution processing module is used to acquire the low-resolution text image to be processed, and to process the low-resolution text image using the super-resolution image generation neural network model to obtain the super-resolution result.

Images