An image high-definition zooming method for a dual-focus camera system

Abstract

The application relates to the field of digital camera image processing, in particular to an image high-definition zooming method for a dual-focus camera system. L Bilinear interpolation is carried out to obtain a zooming result image; in the dynamic zooming mode and the long-focus mode, a super-resolution zooming model is adopted to obtain the zooming result image; in the dual-focus camera system, the digital zooming is carried out in combination with various super-resolution strategies for the overlapped field of view area and the non-overlapped field of view area in the dynamic zooming, so that the zooming precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of digital camera image processing, and in particular to a high-definition image zoom method for a dual-focus camera system. Background Art

[0002] Traditional camera zoom is typically achieved through the movement of optical mechanical components, which limits the camera's size, weight, and manufacturing cost. With the development of computational photography, digital image processing technology has provided a new approach to achieving multi-fold high-definition zoom. This approach can magnify images through related zoom algorithms, overcoming hardware limitations.

[0003] With advances in deep learning and computer vision, super-resolution technology has become an important means of improving image quality. Super-resolution techniques recover details and enhance clarity from low-resolution input images to produce high-resolution images. These methods leverage image statistics and contextual information to improve detail restoration. Furthermore, thanks to advances in hardware technology, mobile cameras now widely use multiple cameras with different focal lengths for different imaging tasks. Therefore, in zoom applications, multiple cameras can be combined for digital zoom, resulting in superior zoom performance. Numerous technologies have been proposed for multi-fold zoom on mobile devices, which can be broadly categorized into three types: single-image-based, multi-frame-based, and reference-image-based super-resolution. Single-image super-resolution typically utilizes internal image information to enhance detail and clarity, while multi-frame super-resolution utilizes multiple low-resolution images and uses image alignment to perform motion estimation to obtain sub-pixel information, which is then used to reconstruct a high-resolution image. Reference-image-based super-resolution utilizes additional high-resolution images, compares and calculates similar information, and fuses this information to produce a high-definition image.

[0004] Current mobile devices typically feature multiple focal length cameras to meet image capture requirements. However, in most devices, each camera operates independently, capturing texture information of varying resolutions through cameras of varying focal lengths, and then processing and transforming the resulting images accordingly. This approach relies heavily on the quality of the images captured by the cameras and fails to fully exploit the complementary information between multiple camera images. Some methods combine image information from two cameras during zooming, but this can lead to image jumps when switching between zoom magnifications, resulting in poor quality zoomed images. Using a reference image for digital zooming can provide realistic texture information. However, global texture information matching transfer has a low tolerance for parallax and therefore performs poorly in complex real-world application scenarios. Local texture information alignment transfer can address the texture transfer problem in complex situations, but the texture information matching accuracy is currently insufficient to achieve high-quality results. Summary of the Invention

[0005] To solve the above problems, the present invention provides a method for high-definition zooming of an image for a dual-focus camera system, which is characterized by comprising the following steps:

[0006] S1. In the target scene, the short focal length camera actively captures image I L And zoom in, then go to step S2;

[0007] S2 determines whether the current zoom magnification is less than the preset magnification x. If so, enter the short focus mode and execute step S3. If not, proceed to step S4.

[0008] S3. Use bilinear interpolation algorithm to process image I L An interpolated image is obtained from the magnified area, and the interpolated image is used as the zoom result image;

[0009] S4. Use a long focal length camera to capture image I in the same target scene H , determine whether the current zoom magnification is greater than the lens magnification, if so, enter the telephoto mode and execute step S6, if not, enter the dynamic zoom mode and execute step S5;

[0010] S5. Image I L Input super-resolution zoom model and take image I H As a reference image, a zoom result image is obtained; the model includes a multi-scale feature extraction module, a similar texture transfer fusion module, a gradient reconstruction module, a feature correction fusion module, and a feature fusion reconstruction module;

[0011] S6. Image I H Input super-resolution zoom model and simultaneously use image I H As the reference image, the zoom result image is obtained.

[0012] Beneficial effects of the present invention:

[0013] This invention divides dual-focus camera image zoom into short-focus mode, dynamic zoom mode, and long-focus mode according to the focal length change factor. Appropriate methods are used for super-resolution in each zoom mode, improving overall zoom image quality. In the dynamic zoom process, this invention does not restrict methods such as multi-scale feature extraction and multi-scale information fusion, allowing for more effective methods to be used.

[0014] In a dual-focus camera system, this invention combines multiple super-resolution strategies to perform digital zoom in overlapping and non-overlapping fields of view during dynamic zoom, improving zoom accuracy. Within the overlapping field of view, multiple scale feature information is acquired from the long- and short-focus images, and texture correlations are calculated for feature matching and transfer. This information is then fused to reconstruct information within the overlapping region. In non-overlapping regions, super-resolution zoom is performed without high-definition reference image information, and low-resolution gradient information is mapped to a high-resolution image via a network to restore high-frequency texture information. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Flow chart of the method of the present invention;

[0016] Figure 2 This is a flowchart of the super-resolution zoom model processing of the present invention;

[0017] Figure 3 Schematic diagram of the structure of the multi-scale feature extraction module of the present invention;

[0018] Figure 4 Schematic diagram of channel feature selection of the present invention;

[0019] Figure 5 This is a structural diagram of the gradient reconstruction module of the present invention;

[0020] Figure 6 This is the structural diagram of the feature correction and fusion module of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] The present invention provides a high-definition zoom method for a dual-focus camera system. Figure 1 As shown, the following steps are included:

[0023] S1. In the target scene, the short-focus camera actively captures the short-focus image and crops the enlarged area of ​​the short-focus image to obtain image I L Then proceed to step S2.

[0024] S2. Determine whether the current zoom magnification is less than the preset magnification x. If so, enter the short focus mode and execute step S3. If not, enter step S4.

[0025] S3. Use bilinear interpolation algorithm to process image I LAn interpolated image is obtained, and the interpolated image is used as the zoom result image.

[0026] Specifically, to achieve efficient large-field-of-view image acquisition, the present invention adopts a dual-focus camera system consisting of a short-focus camera and a long-focus camera, wherein the short-focus camera actively captures images and enters short-focus mode, while the long-focus camera is in an inactive state and does not participate in image capture; the short-focus image captured by the short-focus camera has global information, and during a lower-magnification zoom process, the zoom result map can be obtained through an interpolation algorithm to reduce the system running time.

[0027] In short focus mode, the image I L Use bilinear interpolation algorithm to zoom in and out, as shown below:

[0028] I′ L =f Bilinear (I L )

[0029] Where I′L represents the interpolated image, f Bilinear Represents the bilinear interpolation algorithm.

[0030] S4. Use a long focal length camera to capture image I in the same target scene H , determine whether the current zoom magnification is greater than the lens magnification. If so, enter the telephoto mode and execute step S6. If not, enter the dynamic zoom mode and execute step S5.

[0031] Specifically, the short focal length of the short focal length camera used in the present invention is f S , the telephoto focal length of a telephoto camera is f L , lens magnification = f L / f S .

[0032] S5. Image I L Input super-resolution zoom model and take image I H As a reference image, a zoom result image is obtained; the model includes a multi-scale feature extraction module, a similar texture transfer fusion module, a gradient reconstruction module, a feature correction fusion module, and a feature fusion reconstruction module.

[0033] Specifically, when the current zoom magnification is greater than the preset magnification x and less than the lens magnification n, upsampling the short-focus image actively captured by the short-focus camera will introduce too much smoothing information, and the zoomed short-focus image will be blurry. Based on this, the present invention introduces a dynamic zoom mode to transfer the high-resolution texture information of the long-focus camera to the low-resolution image (i.e., the image I obtained by cropping the magnified area of ​​the short-focus image). L ) corresponds to the texture information position, expressed as:

[0034] ISR =f SR (I L ,I H )

[0035] Among them, f SR Represents the super-resolution zoom model.

[0036] Specifically, step S5 uses the super-resolution zoom model to process the image I L and image I H Get the zoom result image, such as Figure 2 Shown, including:

[0037] S51. Use multi-scale feature extraction module to extract image I L and image I H multi-scale features.

[0038] Specifically, image I L and image I H The image I has the same field of view information, but has a certain scale difference. In order to improve the accuracy of texture information matching, it is necessary to L and image I H Mapping to an approximate scale to match feature information, and improving the information in a single scale by fusing multiple convolution results.

[0039] like Figure 3 As shown, the multi-scale feature extraction module includes a cascaded first convolution block, a second convolution block and a third convolution block; the first convolution block includes a first convolution layer, a second convolution layer, a third convolution layer and a first SA spatial attention layer; the second convolution block includes a first maximum pooling layer, a fourth convolution layer, a fifth convolution layer and a second SA spatial attention layer; the third convolution block includes a second maximum pooling layer, a sixth convolution layer, a seventh convolution layer and a third SA spatial attention layer; step S51 takes image I L or Image I H As the input image, the process of extracting using the multi-scale feature extraction module is as follows:

[0040] S511. Enhance the feature representation capability of the input image through the dimensionality increase operation of the first convolutional layer to obtain the initial feature F0 of the input image; input the initial feature F0 into the second convolutional layer to obtain feature F1, and feature F1 into the third convolutional layer to obtain feature F2; extract deep features through the second and third convolutional layers, respectively processing different feature dimensions and spatial information to ensure that the details and texture information of the image are captured from all angles; it can be expressed as:

[0041] F0=Conv1(I)

[0042] F1=Conv2(F0)

[0043] F2=Conv3(F1)

[0044] Where I represents the input image, Conv1, Conv2, and Conv3 represent the first convolutional layer, the second convolutional layer, and the third convolutional layer, respectively.

[0045] S512. Feature F1 and feature F2 are concatenated and input into the first SA spatial attention layer, and the feature fusion weight is dynamically adjusted to output the first scale feature, which is expressed as:

[0046]

[0047] Wherein, Cat(·) represents the series operation; Represents the first scale feature of the input image I. In the subsequent steps, Represents image I L The i-th scale feature of Represents image I H The i-th scale feature of ; SA(·) represents the SA spatial attention layer.

[0048] S513. To further refine and compress features, the first scale feature is input into the first maximum pooling layer to obtain a downsampled feature F11, the downsampled feature F11 is input into the fourth convolutional layer to obtain a feature F3, and the feature F3 is passed through the fifth convolutional layer to obtain a feature F4;

[0049] S514. Concatenate the features F3 and F4 and input them into the second SA spatial attention layer, outputting the second scale feature.

[0050] S515. Input the second scale feature into the second maximum pooling layer to obtain a downsampled feature F12, input the downsampled feature F12 into the sixth convolutional layer to obtain a feature F5, and the feature F5 is passed through the seventh convolutional layer to obtain a feature F6;

[0051] S516. Concatenate features F5 and F6 and input them into the third SA spatial attention layer, outputting the third scale feature;

[0052] S517. The first scale feature, the second scale feature, and the third scale feature constitute a multi-scale feature of the input image.

[0053] Through the above operations, the module can process and generate feature maps of different scales, thereby forming multi-scale feature information.

[0054] S52. Image I L and image I H The multi-scale features of the texture are input into the similar texture transfer fusion module to obtain the multi-scale transfer features.

[0055] Specifically, step S52 converts the image I L and image I H The multi-scale features of the similar texture transfer fusion module are jointly input to obtain the multi-scale transfer features, including:

[0056] S521. For image I L The i-th scale feature Calculate the significance of each channel feature, sort all channel features of the i-th scale feature in descending order of significance, and select the first m channel features to form a new scale feature; divide each channel feature in the new scale feature into corresponding n×n sub-blocks (n=3, 4, 5, ...), and perform a multiplication of each sub-block from 1 to n from left to right and from top to bottom. 2 After labeling all sub-blocks of the m channel features in the new scale feature, a short-focus block label set is formed; Figure 4 As shown, it can be expressed as:

[0057]

[0058] in, Represents image I L The i-th new scale feature of X(·) represents the significance calculation operation, sort(·) represents the sorting operation, Represents image I L The j-th channel feature of the i-th scale feature is j=1,2,…,M, where M represents the number of channel features; S(·) represents the block operation, Represents image I L The short-focus block label set corresponding to the i-th scale feature, Represents image I L The cth channel in the i-th new scale feature corresponds to the z=1,2,..,n 2 In the present invention, image I L There are three scale features, and each scale feature forms a short-focus block label set according to the above steps.

[0059] S522. For image I H The i-th scale feature Calculate the significance of each channel feature, sort all channel features of the i-th scale feature in descending order of significance, and select the first m channel features to form a new scale feature; divide each channel feature in the new scale feature into n×n sub-blocks, and perform a sparse analysis on each sub-block from 1 to n from left to right and from top to bottom. 2 After labeling all sub-blocks of the m channel features in the new scale feature, a telephoto block label set is formed; it can be expressed as:

[0060]

[0061] in, Represents image I H The i-th new scale feature of Represents image I H The j-th channel feature of the i-th scale feature, Represents image I H The telephoto block label set corresponding to the i-th scale feature, Represents image I H The cth channel in the i-th new scale feature corresponds to the z=1,2,..,n 2 Similarly, in the present invention, image I H There are three scale features, and each scale feature forms a telephoto block label set according to the above steps.

[0062] S523. Perform one-to-one matching of the sub-blocks of the i-th short focus block label set and the i-th long focus block label set according to the labels, calculate the similarity of each matching pair, record the matching pair corresponding to the maximum similarity as the maximum similarity matching pair, and select the matching pair that belongs to image I according to the maximum similarity matching pair. H The labels of the sub-blocks form a position index, which is expressed as:

[0063]

[0064] Among them, s i,c,z represents the similarity between the zth sub-block corresponding to the cth channel in the ith short-focus block label set and the zth sub-block corresponding to the cth channel in the ith long-focus block label set, <, > represents the inner product operation, The short-focus block label set and the long-focus block label set at the same scale are matched one-to-one according to the label to obtain a position index. The present invention has three scales, so three position indexes are calculated.

[0065] S524. Convert the i-th position index into optical flow displacement information, and transform the image I into the image I by deformable convolution according to the optical flow displacement information. L The i-th scale feature and image I H The i-th scale feature is fused to obtain a transfer feature; the image I L and image I H The scale features at the same scale are fused based on the optical flow displacement information corresponding to the same scale to obtain a transfer feature. The present invention has three scales, so three transfer features are calculated.

[0066] S525. Repeat steps S521-S524 to obtain transfer features at three different scales.

[0067] S53. Extract image I through gradient reconstruction module L Multi-gradient features.

[0068] Specifically, in the image captured by the short focal length camera, there is no high-resolution texture information transfer for the area outside the telephoto field of view, so the gradient information itself is used as structural prior information to improve the visual quality of the generated image; the gradient reconstruction module includes 1 initial convolution layer and 3 RTAU blocks, such as Figure 5 As shown; Step S53 extracts image I through the gradient reconstruction module L Multi-gradient features, including:

[0069] S531. Image I L The gradient strength is obtained through the initial convolutional layer and is expressed as:

[0070] I x (x)=I(x+1,y)-I(x-1,y)

[0071] I y (x)=I(x,y+1)-I(x,y-1)

[0072] G0=‖I x (x),I y (x) ‖2

[0073] Among them, I x (x) and I y (x) represents the gradient value in the x direction and y direction respectively, and then the gradient strength G0 is calculated to restore the clear texture of the local area in the image.

[0074] S532. Input the gradient strength G0 into the first RTAU block, combined with image I L The third scale feature Get the first gradient feature G1;

[0075] S533. Input the first gradient feature G1 into the second RTAU block, combined with image I L The second scale feature Get the second gradient feature G2;

[0076] S534. Input the second gradient feature G2 into the third RTAU block, combined with image I L The first scale feature Get the third gradient feature G3;

[0077] S535. The first gradient feature, the second gradient feature and the third gradient feature form image I L Multi-gradient features.

[0078] Specifically, the three RTAU blocks have the same structure, including a convolutional layer, a transformer block, a fusion layer, and an upsampling layer. The processing of the i=1, 2, and 3 RTAU blocks includes:

[0079] A1. Input features are converted into convolutional features through the convolutional layer, and the global structural features of the convolutional features are perceived by the transformer block to obtain global features.

[0080] A2. Convolution features, image I L The scale features and global features are fused in the Fusion layer to obtain scale fusion features;

[0081] A3. Use the upsampling layer to upsample the scale fusion feature to obtain the i-th gradient feature.

[0082] S54. Combined with multi-gradient features, the feature correction fusion module is used to correct the multi-scale transfer features to obtain multi-scale transfer correction features.

[0083] Specifically, directly multi-scale transfer features and image I L After the multi-scale features are fused, deep feature information extraction and reconstruction will be performed, which will result in potential erroneous texture information. Therefore, the present invention combines multi-gradient features to construct a feature correction fusion module to modify and correct the erroneously transferred features to a certain extent.

[0084] The feature correction fusion module includes a splicing layer, a CBAM layer, a convolution layer, and a sigmoid layer. The structure is as follows: Figure 6 As shown, the CBAM layer mainly includes Figure 6 (b) (c) Two attention mechanisms; Step S54 uses the feature correction fusion module to correct the transfer features of the i-th scale to obtain the transfer correction features of the i-th scale, including:

[0085] S541. The i-th gradient feature and the i-th scale transfer feature are passed through a concatenation layer to obtain a concatenated feature, and the concatenated feature is processed using a CBAM layer to obtain an attention enhancement feature;

[0086] S542. Pass the attention enhancement feature through the convolution layer to obtain the attention convolution feature, and pass the i-th gradient feature through the sigmoid layer to obtain the normalized feature;

[0087] S543. Multiply the attention convolution feature and the normalized feature to obtain the transfer correction feature of the i-th scale.

[0088] The above operation can be expressed as:

[0089]

[0090] in, represents the transfer correction feature of the i-th scale, Concat(·) represents the splicing operation, which merges the gradient information and the original features in a specific dimension; CBAM(·) represents the channel and spatial attention mechanism, which is used to enhance key features and suppress irrelevant features; Conv(·) represents the convolution layer operation; sigmod(·) is used to normalize the weight of each channel; Represents element-wise multiplication, which is used to apply the learned weights to the original features to complete feature correction. represents the transfer features of the i=1, 2, and 3 scales.

[0091] S55. Multi-scale transfer correction features and image I L The multi-scale features of the input feature fusion reconstruction module are used to obtain the zoom result image.

[0092] Specifically, step S55 combines the multi-scale transfer correction feature and the image I L The multi-scale feature input feature fusion module obtains the zoom result image, including:

[0093] S551. Transfer the first correction feature and image I L The first scale feature of the image is spliced ​​and then the depth feature is extracted to obtain the first depth feature; the second transfer correction feature and image I L The second scale feature of the image is spliced ​​and then the depth feature is extracted to obtain the second depth feature; the third transfer correction feature and the image I L The third scale feature is spliced ​​together and then the depth feature is extracted to obtain the third depth feature;

[0094] S552. Gradually fuse the three deep features and input them into the convolution layer to obtain the zoom result image.

[0095] S6. Image I H Input the super-resolution zoom model and use itself as the reference image to obtain the zoom result image.

[0096] Specifically, when the current zoom magnification is greater than the lens magnification n, the telephoto mode is entered. At this time, the short focal length camera stops participating in image zooming because it cannot provide sufficient field of view information. The zoom process is expressed as:

[0097] I SR =f SR (I H ,I H )

[0098] At this time, the super-resolution zoom model is still used to achieve higher resolution output.

[0099] In the present invention, unless otherwise clearly stipulated and limited, the terms "installation", "setting", "connection", "fixation", "rotation" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0100] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A high-definition image zoom method for a dual-focus camera system, characterized in that: The following steps are involved: S1. In the target scene, the short-focus camera actively captures the short-focus image and crops the enlarged area of ​​the short-focus image to obtain image I L Then proceed to step S2; S2 determines whether the current zoom magnification is less than the preset magnification x. If so, enter the short focus mode and execute step S3. If not, proceed to step S4. S3. Use bilinear interpolation algorithm to process image I L Obtain an interpolated image, and use the interpolated image as the zoom result image; S4. Use a long focal length camera to capture image I in the same target scene H , determine whether the current zoom magnification is greater than the lens magnification, if so, enter the telephoto mode and execute step S6, if not, enter the dynamic zoom mode and execute step S5; S5. Image I L Input super-resolution zoom model and take image I H As a reference image, a zoom result image is obtained; the model includes a multi-scale feature extraction module, a similar texture transfer fusion module, a gradient reconstruction module, a feature correction fusion module, and a feature fusion reconstruction module; Step S5 includes: S51. Use multi-scale feature extraction module to extract image I L and image I H Multi-scale features; S52. Image I L and image I H The multi-scale features of the textures are input into the similar texture transfer fusion module to obtain the multi-scale transfer features; S53. Extract image I through gradient reconstruction module L Multi-gradient features; S54. Combining multiple gradient features, a feature correction fusion module is used to correct the multi-scale transfer features to obtain multi-scale transfer correction features; S55. Multi-scale transfer correction features and image I L The multi-scale features are input into the feature fusion reconstruction module to obtain the zoom result image; S6. Image I H Input super-resolution zoom model and simultaneously use image I H As the reference image, the zoom result image is obtained.

2. The method for high-definition zooming of a dual-focus camera system according to claim 1, wherein: The multi-scale feature extraction module includes a cascaded first convolution block, a second convolution block and a third convolution block; the first convolution block includes a first convolution layer, a second convolution layer, a third convolution layer and a first SA spatial attention layer; the second convolution block includes a first maximum pooling layer, a fourth convolution layer, a fifth convolution layer and a second SA spatial attention layer; the third convolution block includes a second maximum pooling layer, a sixth convolution layer, a seventh convolution layer and a third SA spatial attention layer; step S51 respectively uses image I L and image I H As the input image, the process of extracting using the multi-scale feature extraction module is as follows: S511. Obtain the initial feature F0 of the input image through the first convolution layer, input the initial feature F0 into the second convolution layer to obtain feature F1, and obtain feature F2 through the third convolution layer; S512. Concatenate the features F1 and F2 and input them into the first SA spatial attention layer, outputting the first scale feature; S513. Input the first scale feature into the first maximum pooling layer to obtain a downsampled feature F11, input the downsampled feature F11 into the fourth convolutional layer to obtain a feature F3, and the feature F3 is passed through the fifth convolutional layer to obtain a feature F4; S514. Concatenate the features F3 and F4 and input them into the second SA spatial attention layer, outputting the second scale feature. S515. Input the second scale feature into the second maximum pooling layer to obtain a downsampled feature F12, input the downsampled feature F12 into the sixth convolutional layer to obtain a feature F5, and the feature F5 is passed through the seventh convolutional layer to obtain a feature F6; S516. Concatenate features F5 and F6 and input them into the third SA spatial attention layer, outputting the third scale feature; S517. The first scale feature, the second scale feature, and the third scale feature constitute a multi-scale feature of the input image.

3. The method for high-definition zooming of a dual-focus camera system according to claim 1, wherein: Step S52: Image I L and image I H The multi-scale features of the similar texture transfer fusion module are jointly input to obtain the multi-scale transfer features, including: S521. For image I L The i-th scale feature, i = 1, 2, 3, calculates the significance of each channel feature, arranges all channel features of the i-th scale feature in descending order of significance, and selects the first m channel features to form a new scale feature; divides the new scale feature into n×n sub-blocks, and performs a sparse ... 2 , forming a short-focus block label set; S522. For image I H The i-th scale feature, i = 1, 2, 3, calculates the significance of each channel feature, arranges all channel features of the i-th scale feature in descending order of significance, and selects the first m channel features to form a new scale feature; divides the new scale feature into n×n sub-blocks, and performs a sparse ... 2 , forming a telephoto block label set; S523. Perform one-to-one matching of the sub-blocks of the i-th short focus block label set and the i-th long focus block label set according to the labels, calculate the similarity of each matching pair to obtain the maximum similarity matching pair, and select the image I from the maximum similarity matching pair. H The labels of the sub-blocks form a position index; S524. Convert the i-th position index into optical flow displacement information, and transform the image I into the image I by deformable convolution according to the optical flow displacement information. L The i-th scale feature and image I H The i-th scale feature is fused to obtain a transfer feature; S525. Repeat steps S521-S524 to obtain transfer features at three different scales.

4. The method for high-definition zooming of a dual-focus camera system according to claim 1, wherein: The gradient reconstruction module includes an initial convolution layer and three RTAU blocks; Step S53 extracts the image I through the gradient reconstruction module L Multi-gradient features, including: S531. Image I L Obtain the gradient strength through the initial convolutional layer; S532. Input the gradient intensity into the first RTAU block, combined with image I L The third scale feature of , the first gradient feature is obtained; S533. Input the first gradient feature into the second RTAU block, combined with image I L The second scale feature of is used to obtain the second gradient feature; S534. Input the second gradient feature into the third RTAU block, combined with image I L The first scale feature of , and the third gradient feature is obtained; S535. The first gradient feature, the second gradient feature and the third gradient feature form image I L Multi-gradient features.

5. The method for high-definition zooming of a dual-focus camera system according to claim 4, wherein: The three RTAU blocks have the same structure, consisting of a convolutional layer, a transformer block, a fusion layer, and an upsampling layer. The processing of the i = 1, 2, and 3 RTAU blocks includes: A1. Input features are converted into convolutional features through the convolutional layer, and the global structural features of the convolutional features are perceived by the transformer block to obtain global features. A2. Convolution features, image I L The scale features and global features are fused in the Fusion layer to obtain scale fusion features; A3. Use the upsampling layer to upsample the scale fusion feature to obtain the i-th gradient feature.

6. The method for high-definition zooming of an image for a dual-focus camera system according to claim 1, wherein: The feature correction fusion module includes a splicing layer, a CBAM layer, a convolution layer, and a sigmoid layer. Step S54 uses the feature correction fusion module to correct the multi-scale transfer features to obtain multi-scale transfer correction features, including: S541. The i-th gradient feature and the i-th scale transfer feature are passed through a concatenation layer to obtain a concatenated feature, and the concatenated feature is processed using a CBAM layer to obtain an attention enhancement feature; S542. Pass the attention enhancement feature through the convolution layer to obtain the attention convolution feature, and pass the i-th gradient feature through the sigmoid layer to obtain the normalized feature; S543. Multiply the attention convolution feature and the normalized feature to obtain the transfer correction feature of the i-th scale.

7. The high-definition image zoom method for a dual-focus camera system according to claim 1, characterized in that: Step S55 combines the multi-scale transfer correction features and image I L The multi-scale feature input feature fusion module obtains the zoom result image, including: S551. The first transfer correction feature and the first scale feature are spliced ​​together and then depth feature extraction is performed to obtain the first depth feature; the second transfer correction feature and the second scale feature are spliced ​​together and then depth feature extraction is performed to obtain the second depth feature; the third transfer correction feature and the third scale feature are spliced ​​together and then depth feature extraction is performed to obtain the third depth feature; S552. Gradually fuse the three deep features and input them into the convolution layer to obtain the zoom result image.

Images