Image translation method and system based on adaptive edge contrast learning and wavelet transform

Abstract

The invention discloses an image translation method and system based on adaptive edge contrast learning and wavelet transform, and the method comprises the steps: coding a to-be-translated image to obtain a content code, a multi-scale high-frequency component and a multi-scale low-frequency component, and carrying out the style extraction and style space enhancement of a reference image to obtain a corresponding style code; respectively preprocessing the multi-scale high-frequency component and the multi-scale low-frequency component to obtain a multi-scale reversal high-frequency component and a multi-scale reversal low-frequency component, and performing frequency enhancement style fusion on the style code, the multi-scale reversal high-frequency component and the multi-scale reversal low-frequency component to obtain a fusion feature; according to the method, intra-class compactness and inter-class difference are enhanced by adopting contrast loss of adaptive edges, a high-quality translated image can be post-synthesized, the translated image can keep domain invariant features of a corresponding source image, and the translated image should be similar to a corresponding reference image in style mode.

Description

Technical Field

[0001] The present invention relates to the technical field of computer image processing, and in particular to an image translation method and system based on adaptive edge contrast learning and wavelet transform. Background Art

[0002] Existing unpaired image-to-image translation can be roughly divided into three categories. The first method learns a bidirectional mapping between the source and target images by forcing the translated image to reconstruct to the source image using a cycle consistency loss during training. The second method is based on disentangled representation, which fully decouples the content and style of the image, while exchanging the content and style vectors in the feature space to achieve image translation. The last method explores contrastive learning to learn the semantic correspondence between input and output. However, if there are mismatched semantic statistics between the source and target domains, the above methods often produce different degrees of semantic distortion. Semantic distortion usually refers to artifacts, hallucinations, and irrelevant attributes that appear in the generated images. This is the result of forced matching (adversarial training) on ​​training datasets with mismatched semantic statistics, and there are three main defects: mode collapse, structural distortion, and style discontinuity. Unpaired images from different visual domains usually have mismatched semantic statistics. Mitigating semantic distortion is a key and challenging issue for improving the fidelity and visual quality of translated images and training downstream tasks using generated data. Summary of the invention

[0003] The purpose of the present invention is to provide an image translation method and system based on adaptive edge contrast learning and wavelet transform to overcome the problems of low fidelity and low visual quality of translated images in the prior art.

[0004] To achieve the above purpose, the technical solution adopted by the present invention is as follows: An image translation method based on adaptive edge contrast learning and wavelet transform comprises the following steps: S1, encode the image to be translated to obtain content encoding, multi-scale high-frequency components and multi-scale low-frequency components, and perform style extraction and style space enhancement on the reference image to obtain the corresponding style encoding; S2, preprocessing the multi-scale high-frequency component and the multi-scale low-frequency component respectively to obtain the multi-scale inverted high-frequency component and the multi-scale inverted low-frequency component, and performing frequency enhancement style fusion on the style coding, the multi-scale inverted high-frequency component and the multi-scale inverted low-frequency component to obtain the fusion feature; S3, decoding the fusion features of the frequency enhanced style fusion to obtain a decoded image, thereby completing the image translation.

[0005] Preferably, a domain-invariant content encoding is extracted from the image to be translated.

[0006] Preferably, Haar wavelet transform is used to decompose the multi-scale coding features into multi-scale high-frequency components and multi-scale low-frequency components.

[0007] Preferably, the obtained multi-scale frequency components are transformed from the frequency domain to the spatial domain using a discrete wavelet inverse transform module, thereby obtaining corresponding multi-scale inverted low frequency and multi-scale inverted high frequency, and then the multi-scale inverted low frequency, multi-scale inverted high frequency and style coding are sent to the frequency enhanced style modulation module for adaptive and efficient fusion.

[0008] Preferably, the multi-scale inverted low frequency and the multi-scale inverted high frequency are subjected to interpolation and convolution operations to generate corresponding low-frequency statistical characteristics and high-frequency statistical characteristics, and the low-frequency scaling parameters and translation parameters are modulated by the high-frequency statistical characteristics to obtain fused frequency scaling parameters and translation parameters respectively, and the fused modulation parameters are used to modulate the feature activation after position normalization.

[0009] Preferably, the AdaIN method is used to activate the feature Modulate the style:

[0010] in, and From the style code The learnable style modulation parameters extracted from and Represent the channel-by-channel mean and standard deviation, represents the final stylized output activation.

[0011] An image translation system based on adaptive edge contrast learning and wavelet transform includes a preprocessing module, a fusion module and a translation module: The preprocessing module encodes the image to be translated to obtain content encoding, multi-scale high-frequency components and multi-scale low-frequency components, and performs style extraction and style space enhancement on the reference image to obtain the corresponding style encoding; A fusion module pre-processes the multi-scale high-frequency component and the multi-scale low-frequency component to obtain a multi-scale inverted high-frequency component and a multi-scale inverted low-frequency component, and performs frequency enhancement style fusion on the style coding, the multi-scale inverted high-frequency component and the multi-scale inverted low-frequency component to obtain a fusion feature; The translation module decodes the fusion features of the frequency enhanced style fusion to obtain a decoded image, thereby completing the translation of the image.

[0012] Preferably, a domain-invariant content encoding is extracted from the image to be translated.

[0013] Preferably, Haar wavelet transform is used to decompose the multi-scale coding features into multi-scale high-frequency components and multi-scale low-frequency components.

[0014] Preferably, the obtained multi-scale frequency components are transformed from the frequency domain to the spatial domain using a discrete wavelet inverse transform module, thereby obtaining corresponding multi-scale inverted low frequency and multi-scale inverted high frequency, and then the multi-scale inverted low frequency, multi-scale inverted high frequency and style coding are sent to the frequency enhanced style modulation module for adaptive and efficient fusion.

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: The invention discloses an image translation method based on adaptive edge contrast learning and wavelet transform, wherein the method encodes an image to be translated to obtain content coding, multi-scale high-frequency components and multi-scale low-frequency components, extracts style and enhances style space of a reference image to obtain corresponding style coding, pre-processes the multi-scale high-frequency components and the multi-scale low-frequency components respectively to obtain multi-scale inverted high-frequency components and multi-scale inverted low-frequency components, performs frequency enhancement style fusion on the style coding, the multi-scale inverted high-frequency components and the multi-scale inverted low-frequency components to obtain fusion features, and adopts adaptive edge contrast loss to enhance intra-class compactness and inter-class difference. The invention can post-synthesize high-quality translated images, the translated images can retain domain-invariant features of their corresponding source images, and the translated images should be similar to their corresponding reference images in terms of style mode.

[0016] Preferably, the present application uses discrete wavelet transform to process the low-frequency and high-frequency information of the input image, effectively integrating the feedforward feature and the reverse frequency information. The present invention can effectively reduce semantic distortion. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is a flow chart of an image translation method based on adaptive edge contrast learning and wavelet transform in an embodiment of the present invention.

[0018] Figure 2 Schematic diagram of the discriminator structure in an embodiment of the present invention.

[0019] Figure 3 Schematic diagram of qualitative comparison of the GTA→Cityscapes task in an embodiment of the present invention.

[0020] Figure 4 Schematic diagram of qualitative comparison of the Horse→Zebra task in an embodiment of the present invention.

[0021] Figure 5 Schematic diagram of the generator framework structure in an embodiment of the present invention.

[0022] Figure 6Schematic diagram of the framework structure of the style modulation module in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] like Figure 1 As shown, the present invention provides an image translation method based on adaptive edge contrast learning and wavelet transform, which can generate identity-preserving and high-fidelity translation images, and specifically includes the following steps: S1, encode the image to be translated to obtain content encoding, multi-scale high-frequency components and multi-scale low-frequency components, and perform style extraction and style space enhancement on the reference image to obtain the corresponding style encoding; S2, preprocessing the multi-scale high-frequency component and the multi-scale low-frequency component respectively to obtain the multi-scale inverted high-frequency component and the multi-scale inverted low-frequency component, and performing frequency enhancement style fusion on the style coding, the multi-scale inverted high-frequency component and the multi-scale inverted low-frequency component to obtain the fusion feature; S3, decoding the fusion features of the frequency enhanced style fusion to obtain a decoded image, thereby completing the image translation.

[0026] In the specific implementation of this application, the translation image Encode to get content encoding , multi-scale high frequency components and multi-scale low-frequency components , and at the same time, the style of the reference image is extracted and the style space is enhanced to obtain the corresponding style encoding .

[0027] The low-frequency component (LL) captures the overall structure and contour of the image, and the high-frequency component (HF = LH+HL+HH) contains subtle edges and fine-grained contour details. Based on this, this application designs a wavelet-guided generator, in which a style modulation module for frequency enhancement is set. The style modulation module adaptively fuses the obtained multi-scale high-frequency components and multi-scale low-frequency components to supplement the source domain identity information and rich details. Under the guidance of the style modulation module, the generator can generate identity-preserving and distortion-free translated images.

[0028] like Figure 5 As shown, the generator includes an encoder and a decoder , encoder Used to extract images to be translated Domain-Invariant Content-Encoding , decoder To generate mixed images , the mixed image can retain the image to be translated identity, while reflecting the reference image style.

[0029] Specifically, the encoder includes several convolutional layers, a normalization layer, an activation layer and four discrete wavelet transform (DWT) modules; the present application adopts a simple and effective Haar wavelet transform to transform the multi-scale coding features ( The index value of the reference layer) is decomposed into high-frequency components and low-frequency components of the corresponding scale ( and ). These frequency components are fed into the FSM module of the decoder through skip connections for subsequent image generation.

[0030] Decoder: In order to achieve image stylization and effectively integrate the inverted frequency features and spatial features, this application adopts the frequency enhanced style modulation (FSM) module and the style modulation based residual block (FSM ResBlk). Based on the proposed style modulation module, such as Figure 6 As shown in FIG. 1 , the decoder of the present application uses 4 FSM ResBlk and 4 inverse discrete wavelet transform (IDWT) modules. The inverse discrete wavelet transform module is used to transform the obtained multi-scale frequency components from the frequency domain to the spatial domain, thereby obtaining the corresponding multi-scale inverted low-frequency and multi-scale inversion of high frequency ( ). Then multi-scale inversion of low frequency , multi-scale inversion of high frequency and style encoding The images are sent to the frequency-enhanced style modulation module for adaptive and efficient fusion. During the decoding process, a skip connection is used to upsample and sum the activation values ​​of the RGB outputs corresponding to different resolutions.

[0031] Specifically, the style modulation module has three external inputs: a multi-scale inverted low-frequency , a multi-scale inversion high frequency and a style code containing the style pattern of the reference image In order to achieve the first goal, that is, to effectively combine the inversion frequency and spatial domain features, the multi-scale inversion low frequency and multi-scale inversion high frequency are interpolated and convolved to generate the corresponding low-frequency statistical characteristics and high-frequency statistical characteristics (learnable high and low frequency modulation parameters: ( , ) and( , )). Through high frequency statistical characteristics ( , ) Modulate the low-frequency scaling parameter and the translation parameter to obtain the fused frequency scaling parameters and translation parameters : (1) in and denote element-wise multiplication and addition, respectively. In addition, all modulation parameters and input activations Keep the same shape, and then use the fused modulation parameters to normalize the position of the normalized feature activation To modulate: (2) At this point, the frequency information of the image to be translated has been added to the feature activation, completing the first goal. Subsequently, the feature activation needs to be stylized. Without damaging the original semantic information, the modulated feature activation is stylized by skip connections. and the original input activation Add; Finally, the AdaIN method is used to activate the features Modulate the style: (3) in, and From the style code The learnable style modulation parameters extracted from and Represent the channel-by-channel mean and standard deviation, represents the final stylized output activation.

[0032] By extending the existing cross-domain distance consistency (CDC), this application adopts distribution statistics regularization (DSR) with the goal of transferring domain-invariant knowledge (identity, structure, and pose) from the source domain to the target domain.

[0033] Recalling the original cross-domain distance consistency (CDC), the relative pairwise distances are maintained before and after adaptation to overcome mode collapse. Specifically, CDC samples a batch of N + 1 noise vectors and use the pairwise similarity in the feature space to build a N Then, by optimizing the KL divergence, the probability distribution of the source generator and the adaptation generator are made consistent.

[0034] Given a source domain N Sample and a reference image , first send them to the generator to get the corresponding response Style translation image , and then use the discriminator's feature extractor 𝐹 to extract the multi-layer representation of the source image (the image to be translated) and the translated image respectively (the total number of layers is L ).in, layer, The probability distribution of samples is given by the pairwise distances between the source and translated images: (4) Where CosSim is the cosine similarity distance, the right superscript 𝑠 represents the source domain, and the right superscript 𝑡 represents the target domain. First, the statistical characteristics (i.e., mean and standard deviation) of the multi-layer probability distribution between the source domain and the target domain are calculated: (5) Then the statistical characteristics of the source domain and the target domain are regularized, which also forms distribution statistical regularization (DSR): (6) In addition, similar to CDC, we use Jensen-Shannon divergence to encourage the target domain to have a similar probability distribution as the source domain, across layers and image instances: (7) Different from the original CDC which uses KL divergence for distribution matching, which may lead to unstable training in the discriminator representation space, this application adopts JS divergence.

[0035] This application adopts a consistency discriminator (CD) to further alleviate overfitting, and proposes a style space enhancement (SSA) to encourage style consistency between the translated image and the reference image.

[0036] The effect of contrastive learning depends largely on the number of negative samples, sampling strategy, and batch size. In this application, Barlow Twins is used, which does not require a large batch size or negative samples, and learns useful representations to avoid model collapse.

[0037] like Figure 2 As shown, the true and false images , and their respective distorted versions ( and ) are input into the feature extractor of the discriminator respectively, and then the Barlow Twins metric is used to represent the pairs (i.e. and , and ) and force these matrices to approach identity. By doing so, the discriminator improves its ability to distinguish between real and fake images and encourages the generator to produce high-fidelity translation images. Experiments show that these two regularizations (i.e., Real Twins and Fake Twins ) can alleviate overfitting. The fundamental reason is that regularization improves the discriminator's ability to distinguish representations, and thus can provide more useful guidance signals for the generator. On the other hand, SSA minimizes the cross-correlation matrix between them (i.e., Style Twins ), ensure twins style encoding ( , ) is not affected by noise in the style space This helps style transfer. Taking Real Twins as an example, we first define and The cross-correlation matrix between : (8) in represents the sample batch, refers to the vector dimension of the input representation, and then Approximating the Identity Matrix To define : . (9) Similarly, by using the cross-correlation matrix For use in , ) and( , ) instead of the corresponding input ( , ) to define and . All inputs ( , ), ( , ) and( , ) have the same shape as the cross-correlation matrix .here represents the feature dimension. is a positive scalar that determines the importance of two terms.

[0038] Example This application conducts qualitative and quantitative comparisons on four 100-sample datasets: AFHQ, Selfie→Anime, Summer↔Winter, and LSUNchurches; among them, the first two datasets require shape deformation, and the latter two datasets require shape preservation.

[0039] Baselines: The proposed method is compared with six existing unpaired image translation models, including Munit, SwapAE, FDIT, Santa, StarGAN v2, and StyleD. For all baselines, the official published codes are used to produce the comparison results without changing their configurations and hyperparameters.

[0040] Datasets: The effectiveness of the proposed method is verified on four datasets that require shape deformation (Animal Faces HQ (AFHQ) and Selfie Anime) or shape retention (Summer Winter Yosemite and LSUN churches). Specifically, 100 training samples are randomly selected for each domain in the original dataset to form a limited data setting. In addition, the models and baselines of this application are trained and tested at 256 × 256 resolution.

[0041] Implementation details: The generator adopts an encoder-decoder structure based on skip connections, and the discriminator adopts a residual network design; this application uses the Adam optimizer, where , , the model was trained in 0.2M batches with a batch size of 8; non-saturated adversarial loss and 𝑅1 regularization were used, and the learning rates of G and D were set to All experiments are performed on 4 NVIDIA3090 GPUs.

[0042] Evaluation metrics: On the one hand, the Fréchet Inception Distance (FID) and the Learned Perceptual Image Patch Similarity (LPIPS) are used as evaluation metrics, which have been widely used in image generation. FID measures the feature distance between the generated distribution and the target distribution, and the smaller the value, the higher the generation quality. LPIPS measures the diversity of the generated images, and the smaller the value, the better the diversity. On the other hand, in order to measure the identity preservation and style similarity of the generated images, this application introduces the Self-Similarity Distance (SSD) and the Single-Image FID (SIFID), the smaller the better. SSD measures the difference in the self-similarity of the keys extracted from the deepest attention module of the transformer between the input and output. Similar to the Gram distance, SwapAE shows and verifies that SIFID is suitable for measuring style similarity by matching the FID between two feature distributions, each of which is generated from a single image.

[0043] Qualitative results: Figure 3 and Figure 4 A visual comparison of four translation tasks is shown. Figure 3 As shown, for the AFHQ and Selfie→Anime datasets, the method of the present application not only produces certain shape deformations (such as animal ears, human eyes), but also obtains satisfactory visual quality. Specifically, the translation results of the method of the present application show strong identity fidelity (for example, animal postures, facial contours) and style consistency similar to the reference images. In contrast, Munit, FDIT, and StarGAN v2 overfit to limited training samples, resulting in significant mode collapse. SwapAE tends to produce outputs close to the input, ignoring the degree of stylization. In the Selfie→Anime task, the translation output of the unilateral translation method Santa is better than other baselines, but there is obvious structural distortion. StyleD also suffers from severe structural distortion, thus sacrificing the identity of the source domain. Figure 4 Presenting Summer The comparison between Winter and LSUN churches is similar to Figure 3 The observations in are roughly consistent. In general, our method produces distortion-free translated images that highly preserve the identity structure of the source domain. The overall shape and local details are well preserved, while the color, lighting, and even weather of the reference image are well transferred.

[0044] Table 1 Comparison results of FID, LPIPS, SSD and SIFID indicators on four 100-sample datasets

[0045] Quantitative comparison: Table 1 shows the comparison of the FID and LPIPS indicators of the four datasets. As can be seen from Table 1, our method obtains the best FID value and the best LPIPS on the two datasets, which shows that our method successfully achieves high-quality and diverse generation under the condition of data scarcity. Since FDIT and StarGAN v2 produce repeated samples, their LPIPS values ​​are close to zero. For the AFHQ and church datasets, StyleD and Munit achieved the best LPIPS values ​​respectively, but we speculate that this is because both of them produce significantly distorted translation images, which may lead to higher diversity. On the other hand, Table 1 also gives the comparison of the SSD and SIFID indicators on the four datasets. As shown in Table 1, our method obtains the best SIFID score, which shows that the method of this application can transfer the style (lighting, color and texture) of the reference image well.

[0046] In summary, the present invention proposes an image translation method based on adaptive edge contrast learning and wavelet transform. First, in order to solve the overfitting problem under limited data conditions, the distribution statistics regularization (DSR) is introduced by extending the existing cross-domain distance consistency (CDC), and a new consistency discriminator (CD) is proposed based on Barlow Twins. Secondly, in order to better maintain the identity attributes during the translation process, the frequency-enhanced style modulation (FSM) is adopted to adaptively and effectively fuse the high- and low-frequency components of the source image. Finally, in order to improve the style consistency between the translated image and the reference image, inspired by Barlow Twins, the style space enhancement (SSA) strategy is adopted. Therefore, the method of the present application can produce high-fidelity translated images with strong identity retention and style consistency. A large number of experiments show that the method of the present application outperforms the leading baselines both qualitatively and quantitatively. This study has taken an important step towards image-to-image translation under limited data to reduce data dependence and facilitate real-world applications.

[0047] According to the actual implementation requirements, it is specifically set, and no specific restrictions are made in this application. It should be noted that in this application, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements that are not clearly listed, or also includes elements inherent to such process, method, article or equipment. In the absence of more restrictions, the elements defined by the sentence "comprise one..." do not exclude the existence of other identical elements in the process, method, article or equipment including the elements.

[0048] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest range consistent with the principles and novel features applied for herein.

Claims

1. An image translation method based on adaptive edge contrast learning and wavelet transform, characterized in that: The following steps are involved: S1, encode the image to be translated to obtain content encoding, multi-scale high-frequency components and multi-scale low-frequency components, and perform style extraction and style space enhancement on the reference image to obtain the corresponding style encoding; S2, preprocessing the multi-scale high-frequency component and the multi-scale low-frequency component respectively to obtain the multi-scale inverted high-frequency component and the multi-scale inverted low-frequency component, and performing frequency enhancement style fusion on the style coding, the multi-scale inverted high-frequency component and the multi-scale inverted low-frequency component to obtain the fusion feature; S3, decoding the fusion features of the frequency enhanced style fusion to obtain a decoded image, thereby completing the image translation.

2. The image translation method based on adaptive edge contrast learning and wavelet transform according to claim 1, characterized in that: Extracting domain-invariant content encodings from images to be translated.

3. The image translation method based on adaptive edge contrast learning and wavelet transform according to claim 1, characterized in that: Haar wavelet transform is used to decompose the multi-scale coding features into multi-scale high-frequency components and multi-scale low-frequency components.

4. The image translation method based on adaptive edge contrast learning and wavelet transform according to claim 1, characterized in that: The obtained multi-scale frequency components are transformed from the frequency domain to the spatial domain using the inverse discrete wavelet transform module to obtain the corresponding multi-scale inverted low frequency and multi-scale inverted high frequency. Then the multi-scale inverted low frequency, multi-scale inverted high frequency and style coding are sent to the frequency enhanced style modulation module for adaptive and efficient fusion.

5. The image translation method based on adaptive edge contrast learning and wavelet transform according to claim 1, characterized in that: The multi-scale inverted low frequency and the multi-scale inverted high frequency are subjected to interpolation and convolution operations to generate corresponding low-frequency statistical characteristics and high-frequency statistical characteristics. The low-frequency scaling parameters and translation parameters are modulated by the high-frequency statistical characteristics to obtain fused frequency scaling parameters and translation parameters respectively. The fused modulation parameters are used to modulate the feature activation after position normalization.

6. The image translation method based on adaptive edge contrast learning and wavelet transform according to claim 5, characterized in that: AdaIN method is used to activate features Modulate the style: in, and From the style code The learnable style modulation parameters extracted from and Represent the channel-by-channel mean and standard deviation, represents the final stylized output activation.

7. An image translation system based on adaptive edge contrast learning and wavelet transform, characterized in that: Including preprocessing module, fusion module and translation module: The preprocessing module encodes the image to be translated to obtain content encoding, multi-scale high-frequency components and multi-scale low-frequency components, and performs style extraction and style space enhancement on the reference image to obtain the corresponding style encoding; A fusion module pre-processes the multi-scale high-frequency component and the multi-scale low-frequency component to obtain a multi-scale inverted high-frequency component and a multi-scale inverted low-frequency component, and performs frequency enhancement style fusion on the style coding, the multi-scale inverted high-frequency component and the multi-scale inverted low-frequency component to obtain a fusion feature; The translation module decodes the fusion features of the frequency enhanced style fusion to obtain a decoded image, thereby completing the translation of the image.

8. The image translation system based on adaptive edge contrast learning and wavelet transform according to claim 7, characterized in that: Extracting domain-invariant content encodings from images to be translated.

9. The image translation system based on adaptive edge contrast learning and wavelet transform according to claim 7, characterized in that: Haar wavelet transform is used to decompose the multi-scale coding features into multi-scale high-frequency components and multi-scale low-frequency components.

10. The image translation system based on adaptive edge contrast learning and wavelet transform according to claim 7, characterized in that: The obtained multi-scale frequency components are transformed from the frequency domain to the spatial domain using the inverse discrete wavelet transform module to obtain the corresponding multi-scale inverted low frequency and multi-scale inverted high frequency. Then the multi-scale inverted low frequency, multi-scale inverted high frequency and style coding are sent to the frequency enhanced style modulation module for adaptive and efficient fusion.

Images