JPEG lossless recompression method based on multi-scale reference

Abstract

The invention provides a JPEG lossless recompression method based on multi-scale reference, and the method comprises the steps: inputting a current JPEG image and a reference JPEG image, carrying out the corresponding processing, finally obtaining the current JPEG coding data, and outputting the current JPEG image, and the corresponding processing comprises the steps of down-sampling, up-sampling, conditional coding, conditional decoding and feature alignment. According to the method, multi-scale condition coding is realized by utilizing the multi-scale correlation characteristic of the reference JPEG and the current JPEG, and the JPEG lossless recompression efficiency and quality are improved.

Description

Technical Field

[0001] The invention relates to an image compression technology, in particular to a JPEG lossless re-compression method based on multi-scale reference. Background Art

[0002] The JPEG (Joint Photographic Experts Group) image compression standard is widely used in the field of digital image processing and storage. It has become one of the most common image compression formats due to its good compression efficiency and low computational complexity. However, with the continuous development of image compression and transmission technology, the traditional JPEG encoding method has gradually exposed its limitations in compression quality and efficiency in certain application scenarios, especially when multiple compression and decompression are required. In the traditional JPEG re-compression process, the image quality often deteriorates significantly, resulting in loss of image details and reduced compression efficiency. Therefore, how to improve the quality and compression efficiency of JPEG images in multiple compression processes has become an important research issue in the field of image processing.

[0003] In recent years, image coding methods based on deep learning have gradually emerged. By learning the super-prior features of images and complex probabilistic models, the compression effect can be significantly improved. In particular, with the support of conditional generation models and feature alignment technology, image recompression can not only achieve a higher compression ratio, but also reduce distortion while retaining more image information. Therefore, studying the JPEG recompression method based on reference images and using the conditional information in the reference image to guide the compression and decompression of the current image has become an effective way to improve the quality of image recompression.

[0004] Most current technologies focus on direct compression optimization by learning image features, but these methods usually ignore the potential of the reference image in the compression process and fail to fully utilize the conditional information of the reference image for effective recompression. Therefore, how to combine the feature information of the reference image and the current image to perform refined recompression through conditional encoding and decoding processes to achieve higher compression quality and efficiency is still a technical problem that needs to be solved urgently. Summary of the invention

[0005] Purpose of the invention: In view of the shortcomings of the prior art, the present invention proposes a JPEG lossless re-compression method based on multi-scale reference, which is particularly suitable for scenarios that need to process large-scale image data, such as cloud storage, video surveillance, image transmission, etc.

[0006] The method of the present invention comprises: inputting a current JPEG image and a reference JPEG image, performing corresponding processing, finally obtaining current JPEG encoded data, and outputting the current JPEG image, wherein the corresponding processing comprises: down sampling, up sampling, conditional encoding, conditional decoding and feature alignment.

[0007] The downsampling includes: taking the input current JPEG image and the reference JPEG image as the first-scale current JPEG image C1 and the first-scale reference JPEG image R1 respectively, wherein C1 is conditionally encoded to obtain the first-scale encoded data B1; downsampling C1 and R1 respectively to obtain the second-scale current JPEG image C2 and the second-scale reference JPEG image R2, wherein C2 is conditionally encoded to obtain the second-scale encoded data B2; downsampling C2 and R2 respectively to obtain the third-scale current JPEG image C3 and the third-scale reference JPEG image R3, wherein C3 is conditionally encoded to obtain the third-scale encoded data B3.

[0008] The upsampling includes: conditionally decoding the third-scale coded data B3 to obtain a third-scale current JPEG image C3, and upsampling the third-scale current JPEG image C3 to obtain an upsampled image C3' of C3; conditionally decoding the second-scale coded data B2 to obtain a second-scale current JPEG image C2, and upsampling the second-scale current JPEG image C2 to obtain an upsampled image C2' of C2; and conditionally decoding the first-scale coded data B1 to obtain a first-scale current JPEG image C1, and using the first-scale current JPEG image C1 as the final output to obtain the current JPEG image.

[0009] The conditional coding comprises the following steps: Step 1-1, extracting super-prior features: parse the input current JPEG image to obtain the DCT coefficients (discrete cosine transform coefficients) y of the three color components, and then pass them through the encoder function of the super-prior encoder Extract super prior features z: ; Perform entropy coding on the extracted super-prior feature z to obtain super-prior feature coding data , the formula is: , in, is the entropy coding function; Step 1-2, conditional information fusion: The goal of conditional information fusion is to combine the super prior feature z with the conditional information c to generate a fused feature representation , through the fusion function Finish: ; When encoding the third-scale current JPEG image C3, the third-scale reference JPEG image R3 is used as condition information c to guide the encoding; when encoding the second-scale current JPEG image C2, the data after feature alignment of the upsampled image C3' and the second-scale reference JPEG image R2 is used as condition information c to guide the encoding; when encoding the first-scale current JPEG image C1, the data after feature alignment of the upsampled image C2' and the first-scale reference JPEG image R1 is used as condition information c to guide the encoding; Step 1-3, conditional probability modeling and entropy coding: represent the fused features Input to the entropy parameter network , entropy parameter network Used to predict the mean of the DCT coefficients y With variance , the formula is: , Then use the mean With variance Probabilistic modeling of the DCT coefficient y is performed, and the DCT coefficient y is set to conform to the Gaussian distribution. The formula is: , in, Indicates that the DCT coefficient y is The probability distribution under the condition, This means that the probability distribution is a distribution with a mean of , the variance is Gaussian distribution of According to the probability distribution of DCT coefficient y, the DCT coefficient y is encoded using the entropy coding algorithm, and finally the encoded data of DCT coefficient y is obtained ; Step 1-4, data packaging and output: encode the data with super-prior features and the encoded data of the DCT coefficients y Integrate to form the final JPEG compressed data packet B: , Among them, Encode is a packaging function.

[0010] The conditional decoding includes: Step 2-1, decapsulation and extraction: Decapsulate the input JPEG compressed data packet B. The formula is: , Among them, Decode is the decapsulation function; Step 2-2, super-prior feature decoding: encode the super-prior feature data After entropy decoding, the decoded super prior feature z is expressed as: , in, represents the entropy decoding function; Step 2-3, conditional information fusion: The decoded super-prior feature z is fused with the input conditional information c to generate the fused feature : , When decoding the third-scale coded data B3, the third-scale reference JPEG image R3 is used as condition information c to guide decoding; when decoding the second-scale coded data B2, the data after feature alignment of the upsampled image C3' and the second-scale reference JPEG image R2 is used as condition information c to guide decoding; when decoding the first-scale coded data B1, the data after feature alignment of the upsampled image C2' and the first-scale reference JPEG image R1 is used as condition information c to guide decoding; Step 2-4, conditional probability modeling and entropy decoding: represent the fused features Input to the entropy parameter network in steps 1-3 , generating the mean With variance , and use the mean With variance The DCT coefficient y is probability modeled, and then according to the probability distribution of the DCT coefficient y, y is decoded using the entropy decoding algorithm to obtain the decoded data of the DCT coefficient y, and finally the current JPEG image is output.

[0011] The feature alignment includes: Step 3-1, feature extraction: perform feature extraction on the reference JPEG image and the current JPEG image respectively, and extract the feature representation of each image through the neural network. Suppose the feature representation of the reference JPEG image is , the feature representation of the current JPEG image is , the process of feature extraction is: , , in, and Represent the reference JPEG image and the current JPEG image respectively, Represents a neural network model for extracting image features; Step 3-2, feature fusion: generate a synthetic feature representation : , in, is the fusion coefficient, which ranges between 0 and 1. This coefficient controls the importance ratio of the reference image and the current image features; Feature Representation Align and obtain the aligned feature representation : ;

[0012] in, It is a transformation function used to align the fused features.

[0013] The present invention also provides an electronic device, comprising a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the described method.

[0014] The present invention also provides a storage medium storing a computer program or instruction. When the computer program or instruction is run on a computer, the steps of the method described are executed.

[0015] Beneficial effects: The present invention realizes reference-based multi-scale conditional coding. By utilizing the correlation characteristics of the reference JPEG and the current JPEG images, the present invention integrates feature information of different scales, performs more accurate probability modeling of DCT coefficients (discrete cosine transform coefficients), improves the compression efficiency of entropy coding, and effectively reduces redundant data. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0017] Figure 1 It is a conditional coding principle diagram.

[0018] Figure 2 This is the conditional decoding schematic.

[0019] Figure 3 It is the feature alignment schematic.

[0020] Figure 4 It is a flowchart for implementing the current JPEG image conditional encoding and decoding.

[0021] Figure 5 It is a comparison chart of the bit saving rate per pixel between the method of the present invention and the traditional image compression scheme.

[0022] Figure 6 This is a comparison chart of bits per pixel between the method of the present invention and traditional image compression schemes. DETAILED DESCRIPTION

[0023] The present disclosure will now be discussed with reference to several example implementations. It should be understood that these implementations are discussed only to enable those skilled in the art to better understand and thus implement the present disclosure, rather than implying any limitation on the scope of the present disclosure.

[0024] An embodiment of the present invention provides a JPEG lossless re-compression method based on multi-scale reference, including: inputting a current JPEG image and a reference JPEG image, performing corresponding processing, finally obtaining current JPEG encoded data, and outputting the current JPEG image, wherein the corresponding processing includes: downsampling, upsampling, conditional encoding, conditional decoding and feature alignment.

[0025] The down sampling includes:

[0026] The input current JPEG image and the reference JPEG image are respectively used as the first-scale current JPEG image C1 and the first-scale reference JPEG image R1, wherein C1 is conditionally encoded to obtain the first-scale encoded data B1; C1 and R1 are respectively downsampled to obtain the second-scale current JPEG image C2 and the second-scale reference JPEG image R2, wherein C2 is conditionally encoded to obtain the second-scale encoded data B2; C2 and R2 are respectively downsampled to obtain the third-scale current JPEG image C3 and the third-scale reference JPEG image R3, wherein C3 is conditionally encoded to obtain the third-scale encoded data B3.

[0027] The upsampling includes: conditionally decoding the third-scale coded data B3 to obtain a third-scale current JPEG image C3, and upsampling the third-scale current JPEG image C3 to obtain an upsampled image C3' of C3; conditionally decoding the second-scale coded data B2 to obtain a second-scale current JPEG image C2, and upsampling the second-scale current JPEG image C2 to obtain an upsampled image C2' of C2; and conditionally decoding the first-scale coded data B1 to obtain a first-scale current JPEG image C1, and using the first-scale current JPEG image C1 as the final output to obtain the current JPEG image.

[0028] The conditional coding comprises the following steps: Step 1-1, extracting super-prior features: parsing the input current JPEG image to obtain DCT coefficients (discrete cosine transform coefficients) y of three color components, and then passing through the encoder function of the super-prior encoder Extract super prior features z: ; Perform entropy coding on the extracted super-prior feature z to obtain super-prior feature coding data , the formula is: , in, is the entropy coding function; Step 1-2, conditional information fusion: The goal of conditional information fusion is to combine the super prior feature z with the conditional information c to generate a fused feature representation , through the fusion function Finish: ; When encoding the third-scale current JPEG image C3, the third-scale reference JPEG image R3 is used as condition information c to guide the encoding; when encoding the second-scale current JPEG image C2, the data after feature alignment of the upsampled image C3' and the second-scale reference JPEG image R2 is used as condition information c to guide the encoding; when encoding the first-scale current JPEG image C1, the data after feature alignment of the upsampled image C2' and the first-scale reference JPEG image R1 is used as condition information c to guide the encoding; Step 1-3, conditional probability modeling and entropy coding: represent the fused features Input to the entropy parameter network , entropy parameter network Used to predict the mean of the DCT coefficients y With variance , the formula is: , Then use the mean With variance Probabilistic modeling of the DCT coefficient y is performed, and the DCT coefficient y is set to conform to the Gaussian distribution. The formula is: , in, Indicates that the DCT coefficient y is The probability distribution under the condition, This means that the probability distribution is a distribution with a mean of , the variance is Gaussian distribution of According to the probability distribution of DCT coefficient y, the DCT coefficient y is encoded using the entropy coding algorithm, and finally the encoded data of DCT coefficient y is obtained ; Step 1-4, data packaging and output: encode the data with super-prior features and the encoded data of the DCT coefficients y Integrate to form the final JPEG compressed data packet B: , Among them, Encode is a packaging function.

[0029] The conditional decoding includes: Step 2-1, decapsulation and extraction: Decapsulate the input JPEG compressed data packet B. The formula is: , Among them, Decode is the decapsulation function; Step 2-2, super-prior feature decoding: encode the super-prior feature data After entropy decoding, the decoded super prior feature z is expressed as: , in, represents the entropy decoding function; Step 2-3, conditional information fusion: The decoded super-prior feature z is fused with the input conditional information c to generate the fused feature : , When decoding the third-scale coded data B3, the third-scale reference JPEG image R3 is used as condition information c to guide decoding; when decoding the second-scale coded data B2, the data after feature alignment of the upsampled image C3' and the second-scale reference JPEG image R2 is used as condition information c to guide decoding; when decoding the first-scale coded data B1, the data after feature alignment of the upsampled image C2' and the first-scale reference JPEG image R1 is used as condition information c to guide decoding; Step 2-4, conditional probability modeling and entropy decoding: represent the fused features Input to the entropy parameter network in steps 1-3 , generating the mean With variance , and use the mean With variance The DCT coefficient y is probability modeled, and then according to the probability distribution of the DCT coefficient y, y is decoded using the entropy decoding algorithm to obtain the decoded data of the DCT coefficient y, and finally the current JPEG image is output.

[0030] The feature alignment includes: Step 3-1, feature extraction: perform feature extraction on the reference JPEG image and the current JPEG image respectively, and extract the feature representation of each image through the neural network. Suppose the feature representation of the reference JPEG image is , the feature representation of the current JPEG image is , the process of feature extraction is: , , in, and Represent the reference JPEG image and the current JPEG image respectively, Represents a neural network model for extracting image features; Step 3-2, feature fusion: generate a synthetic feature representation : , in, is the fusion coefficient, which ranges between 0 and 1. This coefficient controls the importance ratio of the reference image and the current image features; Feature Representation Align and obtain the aligned feature representation : ; in, It is a transformation function used to align the fused features.

[0031] The present invention also provides an electronic device, comprising a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the described method.

[0032] The present invention also provides a storage medium storing a computer program or instruction. When the computer program or instruction is run on a computer, the steps of the method described are executed.

[0033] Figure 1 The implementation principle of conditional coding is shown. First, the input current JPEG image is parsed to obtain the DCT coefficients (discrete cosine transform coefficients) y of the three color components; then, the super-prior encoder is used to extract the super-prior feature z from the DCT coefficient y, where z represents the deep information of the image; then, z is compressed through quantization and entropy coding to generate the coded data of the super-prior feature. ; Then, the extracted super-prior feature z is fused with the input condition information c to obtain the fused feature The fusion process can adjust the relative importance of the reference JPEG image and the current JPEG image features through weighted fusion; then, the fusion result is passed to the entropy parameter network , generates the mean of the DCT coefficients y With variance , which is used for subsequent probability modeling and compression. Using these generated means and variances, probability modeling is performed on the DCT coefficient y, and the DCT coefficient y is compressed by entropy coding technology to obtain the encoded data of the DCT coefficient y ; Finally, by encoding the data with the hyper-prior feature and the encoded data of the DCT coefficients y Integrate to form the final JPEG compressed data packet B.

[0034] Figure 2 The implementation principle of conditional decoding is shown. First, the input JPEG encoded data B is decapsulated to extract the super-prior feature encoded data and the encoded data of the DCT coefficients y ; Then, entropy decoding technology is used to encode the super-prior feature data Decode and restore the super-prior feature z through the super-prior decoder, and fuse the super-prior feature z with the input conditional information c to generate a fused feature ; Then, the fusion features Passed into the entropy parameter network , generating the mean for decoding With variance ; Finally, using the mean With variance The encoded data of the DCT coefficient y is entropy decoded, and then the final current JPEG image is gradually restored.

[0035] Figure 3 The implementation principle of feature alignment is shown. First, feature extraction is performed on the reference JPEG image and the current JPEG image to obtain feature representations of the two images; then, a synthetic feature representation is generated by weighted fusion of the two feature representations. The fusion process adjusts the relative importance of the features of the reference JPEG image and the current JPEG image by setting the fusion coefficient; next, the fused features are aligned using the spatial transformation network to ensure that the features of the reference JPEG image and the current JPEG image are consistent in space, thereby achieving matching and consistency of the JPEG images. This feature alignment process provides more accurate image feature support for subsequent conditional encoding and decoding.

[0036] Figure 4The conditional encoding and decoding process of the current JPEG image is shown. First, the input current JPEG image and the reference JPEG image are used as the first-scale current JPEG image C1 and the first-scale reference JPEG image R1 respectively; C1 and R1 are downsampled to obtain the second-scale current JPEG image C2 and the second-scale reference JPEG image R2 respectively; then C2 and R2 are downsampled to obtain the third-scale current JPEG image C3 and the third-scale reference JPEG image R3 respectively; then, R3 is used as conditional information to guide C3 to conditionally encode to obtain the coded data B3 of C3, and R3 is used as conditional information to guide B3 to conditionally decode and restore C3; then, C3 is upsampled The upsampled image C3' is obtained, and the features of C3' and R2 are aligned. The data after the feature alignment is used to guide C2 to perform conditional encoding to obtain the encoded data B2 of C2, and the data after the feature alignment is used to guide B2 to perform conditional decoding to restore C2; ​​then, C2 is upsampled to obtain the upsampled image C2', and C2' and R1 are feature aligned. The data after the feature alignment is used to guide C1 to perform conditional encoding to obtain the encoded data B1 of C1, and the data after the feature alignment is used to guide B1 to perform conditional decoding to restore C1, and C1 is used as the final output current JPEG image, and the encoded data B1, B2, and B3 are integrated to form the current JPEG image encoded data.

[0037] Figure 5 and Figure 6 The performance indicators of bit saving rate per pixel and bit per pixel are shown to compare the method of the present invention (denoted as Ours) with four traditional single-image compression schemes and a similar image compression algorithm. The schemes are JPEG, Lepton, JPEG XL, Brunsli and the frequency domain block matching algorithm Frequency Domain Block Matching (denoted as FDBM, paper source: Hongwei Sha, Ming Lu, and Zhan Ma, "LosslessJPEG Recompression for Similar Images via Frequency Domain Block Matching," IEEE PCS, Feb. 2024), where FDBM is a similar image compression algorithm. The bar chart shows the performance of the six schemes on the Caltech Pedestrian dataset, UVG ​​1080P video dataset, and UVG 4K dataset. It can be seen that compared with the single-image compression algorithm and the similar image compression algorithm FDBM, the present method can achieve lower bits per pixel and higher bit saving rate per pixel.

[0038] The present invention provides a JPEG lossless recompression method based on multi-scale reference, which adopts multi-scale reference, super-prior feature extraction and compression, conditional encoding and decoding, and feature alignment technologies to achieve efficient image compression and restoration. The present invention can adapt to the image compression requirements of different resolutions, and further improve the compression effect and restoration accuracy. In general, the method of the present invention has high efficiency and wide application potential in the field of image compression and restoration, and is particularly suitable for scenes that require high compression ratios and high image quality requirements, such as image storage and transmission.

Claims

1. A JPEG lossless recompression method based on multi-scale reference, characterized in that: include: The current JPEG image and the reference JPEG image are input, and corresponding processing is performed to finally obtain the current JPEG encoded data, and the current JPEG image is output. The corresponding processing includes: downsampling, upsampling, conditional encoding, conditional decoding and feature alignment.

2. The method according to claim 1, characterized in that The down sampling includes: The input current JPEG image and the reference JPEG image are respectively used as the first-scale current JPEG image C1 and the first-scale reference JPEG image R1, wherein C1 is conditionally encoded to obtain the first-scale encoded data B1; C1 and R1 are respectively downsampled to obtain the second-scale current JPEG image C2 and the second-scale reference JPEG image R2, wherein C2 is conditionally encoded to obtain the second-scale encoded data B2; C2 and R2 are respectively downsampled to obtain the third-scale current JPEG image C3 and the third-scale reference JPEG image R3, wherein C3 is conditionally encoded to obtain the third-scale encoded data B3.

3. The method according to claim 2, characterized in that The upsampling comprises: The third-scale coded data B3 is conditionally decoded to obtain a third-scale current JPEG image C3, and the third-scale current JPEG image C3 is upsampled to obtain an upsampled image C3' of C3; the second-scale coded data B2 is conditionally decoded to obtain a second-scale current JPEG image C2, and the second-scale current JPEG image C2 is upsampled to obtain an upsampled image C2' of C2; the first-scale coded data B1 is conditionally decoded to obtain a first-scale current JPEG image C1, and the first-scale current JPEG image C1 is used as the final output to obtain the current JPEG image.

4. The method according to claim 3, characterized in that The conditional coding comprises the following steps: Step 1-1, extracting super-prior features: parse the input current JPEG image to obtain the DCT coefficients y of the three color components, and then pass the encoder function of the super-prior encoder Extract super prior features z: ； Perform entropy coding on the extracted super-prior feature z to obtain super-prior feature coding data , the formula is: ， in, is the entropy coding function; Step 1-2, conditional information fusion: The goal of conditional information fusion is to combine the super prior feature z with the conditional information c to generate a fused feature representation , through the fusion function Finish: ； When encoding the third-scale current JPEG image C3, the third-scale reference JPEG image R3 is used as condition information c to guide the encoding; when encoding the second-scale current JPEG image C2, the data after feature alignment of the upsampled image C3' and the second-scale reference JPEG image R2 is used as condition information c to guide the encoding; when encoding the first-scale current JPEG image C1, the data after feature alignment of the upsampled image C2' and the first-scale reference JPEG image R1 is used as condition information c to guide the encoding; Step 1-3, conditional probability modeling and entropy coding: represent the fused features Input to the entropy parameter network , entropy parameter network Used to predict the mean of the DCT coefficients y With variance , the formula is: ， Then use the mean With variance Probabilistic modeling of the DCT coefficient y is performed, and the DCT coefficient y is set to conform to the Gaussian distribution. The formula is: ， in, Indicates that the DCT coefficient y is The probability distribution under the condition, This means that the probability distribution is a distribution with a mean of , the variance is Gaussian distribution of According to the probability distribution of DCT coefficient y, the DCT coefficient y is encoded using the entropy coding algorithm, and finally the encoded data of DCT coefficient y is obtained ; Step 1-4, data packaging and output: encode the data with super-prior features and the encoded data of the DCT coefficients y Integrate to form the final JPEG compressed data packet B: ， Among them, Encode is a packaging function.

5. The method according to claim 4, characterized in that The conditional decoding includes: Step 2-1, decapsulation and extraction: Decapsulate the input JPEG compressed data packet B. The formula is: ， Among them, Decode is the decapsulation function; Step 2-2, super-prior feature decoding: encode the super-prior feature data After entropy decoding, the decoded super prior feature z is expressed as: ， in, represents the entropy decoding function; Step 2-3, conditional information fusion: The decoded super-prior feature z is fused with the input conditional information c to generate the fused feature : ， When decoding the third-scale coded data B3, the third-scale reference JPEG image R3 is used as condition information c to guide decoding; when decoding the second-scale coded data B2, the data after feature alignment of the upsampled image C3' and the second-scale reference JPEG image R2 is used as condition information c to guide decoding; when decoding the first-scale coded data B1, the data after feature alignment of the upsampled image C2' and the first-scale reference JPEG image R1 is used as condition information c to guide decoding; Step 2-4, conditional probability modeling and entropy decoding: represent the fused features Input to the entropy parameter network in steps 1-3 , generating the mean With variance , and use the mean With variance The DCT coefficient y is probability modeled, and then according to the probability distribution of the DCT coefficient y, y is decoded using the entropy decoding algorithm to obtain the decoded data of the DCT coefficient y, and finally the current JPEG image is output.

6. The method according to claim 5, characterized in that The feature alignment includes: Step 3-1, feature extraction: perform feature extraction on the reference JPEG image and the current JPEG image respectively, and extract the feature representation of each image through the neural network. Suppose the feature representation of the reference JPEG image is , the feature representation of the current JPEG image is , the process of feature extraction is: ， ， in, and Represent the reference JPEG image and the current JPEG image respectively, Represents a neural network model for extracting image features; Step 3-2, feature fusion: generate a synthetic feature representation : ， in, is the fusion coefficient, which ranges between 0 and 1. This coefficient controls the importance ratio of the reference image and the current image features; Feature Representation Align and obtain the aligned feature representation : ； in, It is a transformation function used to align the fused features.

7. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 6.

8. A storage medium, characterized in that: A computer program or instruction is stored, and when the computer program or instruction is run on a computer, the steps of the method according to any one of claims 1 to 6 are executed.

Images