Method for processing multi-view images, device, equipment, storage medium and program product
Patent Information
- Application Number
- HK32025117218
- Authority / Receiving Office
- HK · HK
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-08-28
Abstract
Description
1. Description of Multi-view Image Processing Method, Apparatus, Device, Storage Medium, and Program Product Technical Field 5 This disclosure relates to the field of light field super-resolution technology, and more specifically, to a multi-view image processing method, apparatus, electronic device, computer-readable storage medium, and computer program product. Background Art Precision engineering has high precision requirements for product quality. High-resolution images can clearly show the details of workpieces, and are therefore frequently used in precision engineering processing, measurement, or inspection. Multi-view images can provide three-dimensional information, thereby improving the accuracy and reliability of measurements. However, due to the cost of image acquisition equipment, high-resolution images, especially high-resolution multi-view images, are lacking in industrial environments. As mentioned above, how to provide high-resolution multi-view images has become an urgent problem to be solved. 15 The information disclosed in the background art section is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention 20 The object of this disclosure is to provide a multi-view image processing method, apparatus, electronic device, readable storage medium, and program product for super-resolution reconstruction of multi-view images. Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part by practice of this disclosure. According to one aspect of this disclosure, a multi-view image processing method is provided, comprising: acquiring a multi-view image to be processed, the multi-view image to be processed including a first image and a second image, the first image including images of at least two views of a target object, the resolution of the first image being less than the resolution of the second image; performing super-resolution reconstruction processing on the first image to obtain a first reconstructed image; performing a first downsampling processing on the second image, and then performing a reverse stretching processing based on the first reconstructed image to obtain a first fused image, the first fused image (HK 30136960 A 2) fusing the second image into the first reconstructed image corresponding to the images of the at least two views; performing feature matching processing on the first image and the second image to obtain a first feature transfer image, the features of the first feature transfer image fusing features of the second image and features of the first image; fusing the first fused image and the first feature transfer image to obtain a target multi-view image, the resolution of the target multi-view image being greater than the resolution of the first image (5). According to one embodiment of this disclosure, the method further includes: performing feature extraction processing on the first image to obtain a first feature map of the first image; and performing feature extraction processing on the second image to obtain a first feature map and a second feature map of the second image.The size of the second feature map of the second image is larger than the size of the first feature map of the second image; performing a first feature matching process on the first image 10 and the second image to obtain a first feature transfer image includes: performing feature matching on the first feature map of the first image and the first feature map of the second image to obtain the similarity between the first feature map of the second image and the first feature map of the first image; performing reverse stretching on the features in the second feature map of the second image according to the similarity, and reconstructing to obtain a feature transfer image, the feature transfer image including the first feature 15 transfer image. According to an embodiment of the present disclosure, the size of the first feature map of the first image is the same as the size of the first feature map of the second image; performing feature matching on the first feature map of the first image and the first feature map of the second image to obtain the similarity between the first feature map of the second image and the first feature map of the first image includes: dividing the first feature map 20 of the first image into a first number of blocks; dividing the first feature map of the second image into the first number of blocks; calculating the similarity between the first feature map of the first image and the first feature map of the second image block by block to obtain a similarity matrix between the first feature map of the first image and the first feature map of the second image. According to one embodiment of this disclosure, the second feature map of the second image includes at least two feature maps of different sizes; the feature transfer image further includes a second feature transfer image, the size of which is larger than the size of the first feature transfer image; fusing the first fused image and the first feature transfer image to obtain a target multi-view image includes: fusing the first fused image and the first feature transfer image to obtain a first feature transfer fused image; performing super-resolution reconstruction processing on the first feature transfer fused image to obtain a second reconstructed image; performing a second downsampling processing on the second image, and then performing reverse stretching processing based on the second reconstructed image to obtain a second fused image; fusing the second fused image and the second feature transfer image to obtain a target multi-view image. According to one embodiment of this disclosure, feature extraction processing is performed on the first image to obtain a first feature map of the first image, including: performing feature extraction processing on the first image using a first encoder to obtain the first feature map of the first image, wherein the first encoder includes a convolutional layer and residual blocks; feature extraction processing is performed on the second image to obtain a first feature map and a second feature map of the second image, including: performing feature extraction processing on the second image using at least one encoding module of a second encoder to obtain a second feature map of the second image; and performing feature extraction processing on the second image using at least two encoding modules of the second encoder.A first feature map of the second image is obtained, and the encoding module of the second encoder includes a convolutional layer and a residual block. According to one embodiment of this disclosure, super-resolution reconstruction processing is performed on the first image to obtain a first reconstructed image, including: extracting features from the first image using a spatially-angle-separable convolution method to obtain a first low-resolution feature map; reconstructing the upsampled first low-resolution feature map using a convolutional layer to obtain a first reconstruction residual map; and merging the first reconstruction residual map with the upsampled first image to obtain the first reconstructed image. According to one embodiment of this disclosure, the first image includes images from at least three perspectives of the target object, including a central perspective; the second image includes an image from the central perspective of the target object. According to another aspect of this disclosure, a multi-view image processing apparatus is provided, comprising: an acquisition module for acquiring a multi-view image to be processed, the multi-view image to be processed including a first image and a second image, the first image including images of at least two views of a target object, the resolution of the first image being less than the resolution of the second image; a super-resolution reconstruction module for performing super-resolution reconstruction processing on the first image to obtain a first reconstructed image; a reverse stretching module for performing a first downsampling processing on the second image and then performing reverse stretching processing based on the first reconstructed image to obtain a first fused image, the first fused image fusing the second image into the first reconstructed image corresponding to the images of the at least two views; a feature matching module for performing feature matching processing on the first image and the second image to obtain a first feature transfer image, the features of the first feature transfer image fusing features of the second image and features of the first image; and a fusion module for fusing the first fused image and the first feature transfer image to obtain a target multi-view image, the resolution of the target multi-view image being greater than the resolution of the first image. According to an embodiment of this disclosure, the apparatus further includes: a first feature extraction module, configured to perform feature extraction processing on the first image to obtain a first feature map of the first image; a second feature extraction module, configured to perform feature extraction processing on the second image to obtain a first feature map and a second feature map of the second image, wherein the size of the second feature map of the second image is larger than the size of the first feature map of the second image; the feature matching module is further configured to perform feature matching on the first feature map of the first image and the first feature map of the second image to obtain a similarity between the first feature map of the second image and the first feature map of the first image; and to perform reverse stretching processing on the features in the second feature map of the second image based on the similarity, and reconstruct to obtain a feature transfer image.The feature transfer image includes the first feature transfer image. According to an embodiment of this disclosure, the feature matching module is further configured to: divide the first feature map of the first image into a first number of blocks; divide the first feature map of the second image into the first number of blocks; calculate the similarity between the first feature map of the first image and the first feature map of the second image block by block, to obtain a similarity matrix between the first feature map of the first image and the first feature map of the second image. According to another aspect of this disclosure, an electronic device is provided, including: a memory, a processor, and executable instructions stored in the memory and executable in the processor, wherein the processor executes the executable instructions to implement any of the methods described above. According to another aspect of this disclosure, a computer-readable storage medium is provided, on which computer-executable instructions are stored, wherein the executable instructions, when executed by a processor, implement any of the methods described above. Embodiments of this disclosure provide a computer program product or computer program, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above. The multi-view image processing method provided in the embodiments of this disclosure acquires a multi-view image including a first image (30 HK 30136960 A 5) and a second image. The first image includes images of at least two views of a target object, and the resolution of the first image is lower than that of the second image. A first reconstructed image is obtained by performing super-resolution reconstruction processing on the first image. After performing a first downsampling process on the second image, a reverse stretching process is performed based on the first reconstructed image to obtain a first fused image that integrates the second image into the first reconstructed image corresponding to the images of at least two views. Then, feature matching processing is performed on the first image and the second image to obtain a first feature transfer image that integrates features from both the second and first images. Finally, the first fused image and the first feature transfer image are fused to obtain a target multi-view image with a resolution greater than that of the first image, thereby achieving super-resolution reconstruction of multi-view images. It should be understood that the above general description and the following detailed description are merely exemplary and do not limit this disclosure. The accompanying drawings describe exemplary embodiments in detail with reference to the accompanying drawings.The above and other objectives, features, and advantages of this disclosure will become more apparent. Figure 1 shows a schematic diagram of a system architecture according to an embodiment of this disclosure. Figure 2 shows a flowchart of a multi-view image processing method according to an embodiment of this disclosure. Figure 3 shows a schematic diagram of the processing procedure of step S204 shown in Figure 2 in one embodiment. Figure 4 shows a schematic diagram of the architecture and feature extraction process of an exemplary feature extraction module. Figure 5 shows a schematic diagram of the architecture and related processing procedure of an exemplary reverse stretching module according to Figures 2 to 4. Figure 6 shows a schematic diagram of the processing procedure of step S208 shown in Figure 2 in one embodiment. Figure 7 shows a schematic diagram of the related processing procedure of an exemplary feature matching according to Figure 6. Figure 8 shows a schematic diagram of the processing procedure of steps S602 and S604 shown in Figure 6 in one embodiment. Figure 9 shows a schematic diagram of the processing procedure of step S606 shown in Figure 6 in one embodiment. Figure 10 shows a schematic diagram of the processing procedure of step S210 shown in Figure 2 in one embodiment. Figure 11 illustrates an overall flow diagram of a hybrid multi-resolution cascaded model 30 HK 30136960 A 6, exemplarily shown according to Figures 2 to 10. Figure 12 shows a block diagram of a multi-view image processing apparatus according to an embodiment of the present disclosure. Figure 13 shows a block diagram of another multi-view image processing apparatus according to an embodiment of the present disclosure. Figure 14 shows a structural schematic diagram of an electronic device according to an embodiment of the present disclosure. 5 Detailed Description Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the present disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The drawings are merely illustrative of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of the present disclosure. However, those skilled in the art will recognize that the technical solutions disclosed herein can be practiced with one or more of the specific details omitted, or other methods, apparatuses, steps, etc., can be employed. In other cases, well-known structures, methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring aspects of this disclosure. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus,Features marked with "first" or "second" may explicitly or implicitly include one or more of those features. In the description of this disclosure, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. The symbol " / " generally indicates that the preceding and following objects are in an "or" relationship. In this disclosure, unless otherwise explicitly specified and limited, terms such as "connection" should be interpreted broadly, for example, it can mean electrical connection or mutual communication; it can mean direct connection or indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this disclosure according to the specific circumstances. In modern industry, precision engineering requires the highest level of precision. Its core lies in achieving machining, measurement, and assembly accuracy at the micrometer or even nanometer level. Compared with general engineering, precision engineering not only requires higher dimensional accuracy and surface quality but also involves complex microstructures and functional characteristics. Machine vision can quickly acquire image information of objects in a non-contact manner and provide real-time quality monitoring and response for production processes in precision engineering. It plays a crucial role in dimensional measurement, online monitoring, and quality inspection in precision engineering. However, its performance relies on using expensive cameras to capture high-resolution images containing information about the state of the measured objects, such as tools, products, and workpieces. Lower resolution limits the clear rendering of image details, making it difficult to accurately identify and analyze subtle defects and complex texture features. Therefore, the quality of the captured industrial images has a significant impact on the performance of machine vision tasks in precision engineering systems. The continuous development of machine vision technology is essential, and recently, significant progress has been made in super-resolution techniques for generating high-quality images from low-resolution images. Deep learning methods, including convolutional neural networks, generative adversarial networks, and diffusion models, have been applied to super-resolution tasks on public datasets. However, due to the lack of industrial image datasets available for super-resolution, research on industrial image super-resolution, especially on images of metallic objects, is relatively limited. Commonly used super-resolution datasets include people, plants, and animals, whose characteristics differ from those of metallic materials. Metallic surfaces are typically highly reflective, which interferes with image detail and increases the difficulty of super-resolution reconstruction. Furthermore, multi-view images are increasingly used in precision engineering. By acquiring image information of an object from different angles, multi-view images can comprehensively present the object's three-dimensional structure and surface features. This provides more comprehensive and accurate information for the detection, measurement, and quality control of complex-shaped objects. This multi-view image acquisition method plays an irreplaceable role in improving detection accuracy and reliability in precision engineering. However, a similar problem exists...The effectiveness of multi-view images in precision engineering remains limited by resolution. Current research on super-resolution mainly focuses on single-image super-resolution. Deep learning-based single-image super-resolution methods pair low-resolution images with corresponding high-resolution images, then use these pairings to train a model that learns the mapping relationship between low-resolution and high-resolution images. Compared to single-image super-resolution, multi-view image super-resolution requires extracting information from multiple images, significantly increasing the research difficulty. Therefore, research on multi-image super-resolution is relatively less extensive than that on single-image super-resolution. Multi-view metal images can present comprehensive key information such as the appearance features and surface defects of metal workpieces. In precision engineering, research on multi-view metal image super-resolution has multiple important benefits. Multi-view metal image super-resolution technology helps to more clearly display subtle defects on metal surfaces, thereby improving the accuracy of defect detection. In addition, multi-view images can provide three-dimensional information about objects. Super-resolution technology can further improve the resolution of these images to reconstruct more accurate three-dimensional models and achieve precise measurement of complex shapes. However, research on super-resolution of multi-view metal images is still insufficient in the field of precision engineering. Therefore, this disclosure provides a multi-view image processing method that employs a reference image-based super-resolution technique. By introducing a high-resolution related image from one viewpoint, its details are fused with the reconstruction results of low-resolution multi-view images based on light field image super-resolution technology, thereby generating a high-resolution multi-view image. This method can achieve super-resolution reconstruction of multi-view metal images, effectively improving the quality of multi-angle metal images and thus enhancing the performance of machine vision tasks in precision engineering. Figure 1 shows an exemplary system architecture 10 to which the multi-view image processing method or multi-view image processing device of this disclosure can be applied. As shown in Figure 1, the system architecture 10 may include terminal devices (1022, 1024), a network 104, and a server 106. Terminal device 1022 may be various electronic devices with a display screen and supporting input and output, including but not limited to smartphones, tablets, laptops, desktop computers, wearable devices, virtual reality devices, smart home devices, etc.; terminal device 1024 may be a device with a shooting function that digitizes the captured image. Terminal device 1022 and terminal device 1024 can be connected via wired or wireless means. Network 104 serves as the medium for providing a communication link between the terminal devices (1022, 1024) and server 106. Network 10420 can include various connection types, such as wired, wireless communication links, or fiber optic cables, etc. Server 106 can be a server or server cluster providing various services, etc.For example, processing servers, database servers, etc. Users can use terminal device 1022 to interact with server 106 via network 104 to receive or send data. For example, a user operates on the software interface of terminal device 1022 to send multi-view images captured by terminal device 1024 to server 106 for processing via network 104. Another example is that a user operates on the software interface of terminal device 1022 to receive the target multi-view image processed by server 106 via network 104 and display the target multi-view image on the software interface of terminal device 1022. It should be understood that the number of terminal devices, networks, and servers in Figure 1 is merely illustrative. 30 HK 30136960 A 9 Depending on implementation needs, there can be any number of terminal devices, networks, and servers. Figure 2 is a flowchart illustrating a multi-view image processing method according to an exemplary embodiment. The method shown in Figure 2 can be applied, for example, to the server side of the system shown in Figure 1, or to the terminal devices of the system shown in Figure 1. Referring to Figure 2, the method 20 provided in this embodiment may include the following steps S202 to S210. In step S202, a multi-view image to be processed is acquired. The multi-view image to be processed includes a first image and a second image. The first image includes images of at least two views of a target object, and the resolution of the first image is lower than the resolution of the second image. In some embodiments, the first image may include images of at least three views of the target object, and the at least three views may include a central view. The first image may be obtained by imaging the target object from different views using a low-resolution imaging device. The target object may be, for example, a metallic object. In some embodiments, the second image may include an image of the target object from the central view. The second image may be obtained by imaging the target object from the central view using a high-resolution imaging device. The view of the central view in the second image may be the same as or slightly different from the view of the central view in the first image. In step S204, super-resolution reconstruction processing is performed on the first image to obtain a first reconstructed image. In some embodiments, a spatially-angle-separable convolution method can be used to extract 20 features from the first image, followed by upsampling. Then, the image is reconstructed through a convolutional layer and merged with the upsampled first image to obtain a first reconstructed image. Exemplary implementations can be seen in Figures 3 and 4. In other embodiments, other light field super-resolution techniques can be used to perform super-resolution reconstruction of the first image, improving the quality of the first image containing spatial and angular information. In step S206, after performing a first downsampling process on the second image, a reverse stretching process is performed based on the first reconstructed image.To obtain a first fused image, the first fused image fuses the second image into a first reconstructed image corresponding to at least two viewpoints. In some embodiments, the image after the second image undergoes a first downsampling process can be back-stretched based on the first reconstructed image and then merged with the first reconstructed image to generate a first fused image, thereby propagating the information of the high-resolution center view of the second image to each viewpoint. For example, the network uses differentiable bicubic interpolation (DBI) for back-stretching, which allows the back-stretching module to be trained end-to-end. An exemplary embodiment can be seen with reference to FIG5. In step S208, feature matching processing is performed on the first image and the second image to obtain a first feature transfer image, the features of which are fused with the features of the second image and the features of the first image. In some embodiments, feature extraction processing can be performed on the first image to obtain a first feature map of the first image, and feature extraction processing can be performed on the second image to obtain a first feature map of the second image and a second feature map with a size larger than the first feature map. Then, the similarity between the first feature map of the second image and the first feature map of the first image is obtained. After upsampling the first feature map of the first image, the features in the second feature map of the second image are reverse stretched according to the similarity, and a feature-transferred image including the first feature-transferred image is reconstructed. An exemplary embodiment can be referred to Figure 6. In step S210, the first fused image and the first feature-transferred image are fused to obtain a target multi-view image, the resolution of which is greater than the resolution of the first image. In some embodiments, the second feature map of the second image may include at least two feature maps of different sizes, all of which can be obtained using the method shown in Figure 8. Correspondingly, the methods of Figure 6 and / or Figure 7 can be used to obtain at least two feature-transferred images, such as the first feature-transferred image and a second feature-transferred image with a size larger than the first feature-transferred image. A cascaded model can be established to fuse 20 feature-transfer images of different sizes into a low-resolution image. For example, a first fused image and a first feature-transfer image can be fused to obtain a first feature-transfer fused image. Then, the first feature-transfer fused image is used as input for super-resolution reconstruction to obtain a second reconstructed image. After performing a second downsampling process on the second image, a reverse stretching process is performed based on the second reconstructed image to obtain a second fused image. Finally, the second fused image and the second feature-transfer image are fused to obtain a multi-view image of the target. Exemplary implementations can be seen in Figures 10 and 11. In some embodiments,The network can be trained using a publicly available multi-view metal object image dataset for 6D pose estimation. Nine views can be selected from a total of 13 views to form a 3×3 multi-view image matrix. Image enhancement operations, including cropping, rotation, and flipping, can be performed on the training set to improve the image quality. According to the multi-view image processing method provided in this disclosure, a first reconstructed image is obtained by performing super-resolution reconstruction processing on a multi-view, low-resolution first image. After performing a first downsampling process on a high-resolution second image, a reverse stretching process is performed based on the first reconstructed image to obtain a first fused image that fuses the second image into the first reconstructed image corresponding to at least two views. Then, feature matching processing is performed on the first and second images to obtain a first feature transfer image that fuses the features of the second image and the features of the first image. Finally, the first fused image and the first feature transfer image are fused to obtain a target multi-view image with a resolution greater than that of the first image. This hybrid input light field super-resolution method combines reference-based super-resolution technology with light field image super-resolution technology, using a high-resolution center view image and a low-resolution light field image as input to improve the resolution of the light field image. This deep learning-based method can solve the problem of multi-view metal image super-resolution and improve the resolution and detail of metal images, filling the gap in current research on multi-view metal image super-resolution. This method can significantly improve the accuracy of metal surface defect detection and the measurement accuracy of complex-shaped metal parts, and can provide useful reference and guidance for the application of super-resolution technology in other fields. Figure 3 shows a schematic diagram of the processing of step S204 shown in Figure 2 in one embodiment. As shown in Figure 3, in this embodiment of the present disclosure, step S204 may further include the following steps S302 to S306. In step S302, a spatial-angle separable convolution method is used to extract features from the first image to obtain a first low-resolution feature map. In some embodiments, efficient feature extraction can be achieved by alternately using two-dimensional convolutions of spatial and angular dimensions. After feature extraction is completed, the network will fuse features from different angles. Fusion is achieved by first combining feature maps by channel, and then further extracting features through deeper convolutional layers. Figure 4 shows a schematic diagram of the architecture and feature extraction process of an exemplary feature extraction module. As shown in Figure 4, the input multi-view low-resolution (LR) image 4002 can be represented as R∈ℝ×××, where × represents the spatial resolution.× represents angular resolution. First, spatial convolution module 402 obtains am*an feature maps of size ah*aw. Then, reshaping module 404 reshapes them from spatial to angular into aw*ah feature maps of size am*an. Next, angular convolution module 406 convolves each am*an block, outputting aw*ah angular fusion feature maps of size am*an. Finally, reshaping module 408 reshapes them from angular to spatial into am*an feature maps of size aw*ah and outputs 4004. Step S304: The upsampled first low-resolution feature map is reconstructed using convolutional layers to obtain a first reconstruction residual map. In some embodiments, the upsampled first low-resolution feature map can be reconstructed using convolutional layers to obtain the first reconstruction residual map. For example, the network can use an efficient 5-PixelShuffle layer to upsample the low-resolution feature map to a high resolution. Step S306: The first reconstructed residual map is merged with the upsampled first image to obtain the first reconstructed image. In some embodiments, the first reconstructed image can be obtained by adding the first reconstructed residual map with the upsampled (e.g., bicubic interpolation) low-resolution image. 10 According to the method provided in the embodiments of this disclosure, a first low-resolution feature map is obtained by extracting features from the first image through a spatially-angle-separable convolutional network. Then, the first reconstructed residual map is reconstructed through a convolutional layer. The first reconstructed residual map is then merged with the upsampled first image. In the super-resolution reconstruction process, angular resolution information is fully utilized, thereby achieving full utilization of complementary information from low-resolution images at different viewpoints.15 Figure 5 shows a schematic diagram of the architecture and related processing of an exemplary reverse stretching module according to Figures 2 to 4. As shown in Figure 5, after feature extraction is performed on the input multi-view LR image 5002 in step S502 (see Figure 4 for an exemplary implementation), it undergoes upsampling processing in step S504 and is then reconstructed through a convolutional layer to obtain a first reconstruction residual image. The first reconstruction residual image is then merged with the low-resolution image after upsampling processing (S506) (S508) to obtain a first reconstructed image. In step S512, after downsampling processing is performed on the center-view HR image 5004 (e.g., the resulting image size is twice that of the multi-view LR image 5002), it is inversely stretched based on the first reconstructed image (S510) and then merged with the first reconstructed image.A first fused image is generated. Then, in step S514, the first fused image is fused with the first feature transfer image obtained in step S208 to generate a high-resolution multi-view image (e.g., magnified by two times). According to the method provided in the embodiments of this disclosure, the complementary information of multi-view low-resolution images based on light field super-resolution technology and the texture features extracted from images of different resolutions based on reference image super-resolution technology are fused. By introducing mixed resolution input, a more realistic multi-view image can be generated from a high-resolution center view input, which can enhance the real texture of metal images. 30 HK 30136960 A 13 Because metal images have less or almost no texture compared to other images. Adding a high-resolution center view image to the network to provide high-frequency information helps to reproduce image details. Using the center view to align images from other viewpoints is computationally efficient and easy to set up using a shooting device. Figure 6 shows a schematic diagram of the processing of step S208 shown in Figure 2 in one embodiment. As shown in Figure 6, in the embodiments of this disclosure, the above step S208 may further include the following steps S606 to S608. In some embodiments, steps S602 to S604 may be performed before step S606. Step S602 involves performing feature extraction processing on the first image to obtain a first feature map of the first image. In some embodiments, feature extraction processing on the first image can be performed using a first encoder including convolutional layers and residual blocks to obtain the first feature map of the first image; an exemplary embodiment can refer to FIG8. Step S604 involves performing feature extraction processing on the second image to obtain a first feature map and a second feature map of the second image, wherein the size of the second feature map of the second image is larger than the size of the first feature map of the second image. In some embodiments, the size of the first feature map of the first image is the same as the size of the first feature map of the second image. In some embodiments, feature extraction processing on the second image can be performed using at least one encoding module of the second encoder to obtain the second feature map of the second image, and feature extraction processing on the second image can be performed using at least two encoding modules of the second encoder to obtain the first feature map of the second image; an exemplary embodiment can refer to FIG8. Step S606 involves performing feature matching on the first feature map of the first image and the first feature map of the second image to obtain the similarity between the first feature map of the second image and the first feature map of the first image. 25 In some embodiments, the first feature map of the first image and the first feature map of the second image can be divided into the same number of blocks, and then the similarity between the first feature map of the first image and the first feature map of the second image can be calculated block by block.Obtain a similarity matrix between the first feature map of the first image and the first feature map of the second image. An exemplary implementation can be seen in Figure 9. In other embodiments, geometric matching, optical flow matching, or other feature matching methods, or image matching methods, can be used to match the features of the second image with the features of the first image. Step S608: Based on the similarity, the features in the second feature map of the second image are reverse stretched, and a feature transfer image is reconstructed, including the first feature transfer image. In some embodiments, after feature matching of the first feature map of the first image and the first feature map of the second image, upsampling is performed, and sub-pixel convolution is used to upsample the low-resolution feature map to a high-resolution image. Then, based on the similarity, the features in the second feature map of the second image are reverse stretched, and the enhanced feature map is reconstructed into a high-resolution image through multiple convolutional layers to obtain the feature transfer image, thereby distorting the details of the high-resolution center view image onto the low-resolution image of each view. 10. In some embodiments, using a second feature map in FIG9 with a size of ×8, ×4, or ×2, a feature transfer image with a size of ×8, ×4, or ×2 can be correspondingly obtained. The first feature transfer image can be, for example, a feature transfer image with a size of ×2. FIG7 illustrates a schematic diagram of an exemplary feature matching correlation process according to FIG6. As shown in Figure 15.7 (for example, using the embodiment shown in Figure 8), after obtaining the first feature map 7002 of the multi-view LR image, the first feature map 7004 of the center-view HR image, and the second feature map 7006 of the center-view HR image (exemplarily shown in Figure 7 with a size of ×2), feature matching (S702) is performed on the first feature map 7002 of the multi-view LR image and the first feature map 7004 of the center-view HR image, followed by upsampling (S704). Then, based on the similarity obtained from feature matching, the features in the second feature map of the second image are reverse stretched based on the upsampled feature map (S706), and then reconstructed (S708) into a high-resolution feature transfer image (e.g., the first feature transfer image 7008). Figure 8 shows a schematic diagram of the processing procedure of steps S602 and S604 shown in Figure 6 in one embodiment. As shown in Figure 8, in this embodiment of the present disclosure, step S602 may further include step S802, and step S604 may further include steps S804 to S806. Step S802 involves performing feature extraction processing on the first image using the first encoder to obtain a first feature map of the first image.The first encoder includes a convolutional layer and residual blocks. In some embodiments, the first encoder may include one convolutional layer and three residual blocks. Step S804: The second image is processed by feature extraction through at least one encoding module of the second encoder to obtain a second feature map of the second image. Step S806: The second image is processed by feature extraction through at least two encoding modules of the second encoder to obtain a first feature map of the second image. The encoding modules of the second encoder include convolutional layers and residual blocks. In some embodiments, the encoding module of the second encoder may include one convolutional layer and three residual blocks, and the size of the output feature map can be reduced by adjusting the stride of the convolutional layer. For example, taking a stride of 2 for the convolutional layer of the encoding module, after the second image passes through the first encoding module, a feature map of size ×8 is obtained; after passing through the second encoding module, a feature map of size ×4 is obtained; after passing through the third encoding module, a feature map of size ×2 is obtained; and after passing through the fourth encoding module, a feature map of size ×1 is obtained. If the size of the first feature map of the first image is ×1, then the second image can be processed by one, two, or three encoding modules of the second encoder to obtain a second feature map of size ×8, ×4, or ×2; the second image can be processed by four encoding modules of the second encoder to obtain a first feature map of size ×1. Figure 9 shows a schematic diagram of the processing of step S606 shown in Figure 6 in one embodiment. As shown in Figure 9, in this embodiment of the present disclosure, step S606 may further include the following steps: Step S902, dividing the first feature map of the first image into a first number of blocks. Step S904, dividing the first feature map of the second image into a first number of blocks. Step S906, calculating the similarity between the first feature map of the first image and the first feature map of the second image block by block to obtain a similarity matrix between the first feature map of the first image and the first feature map of the second image. In some embodiments, taking the second image as the central viewpoint image as an example, each block of the low-resolution image of each view in the first image is compared 25 times with each block of the high-resolution image of the central view, so that each view image is assigned a similarity map, which records its block-by-block similarity score relative to the central view image. The second feature map of the second image can then be attentionally transferred to the feature map of the first image based on this similarity map. According to the method provided in embodiments of this disclosure, a similarity matrix is calculated by matching feature map blocks of the central view image and other view images one by one, and by calculating the similarity between the feature maps...30 HK 30136960 A 16 Regions with similar textures are found, enabling feature matching operations between the first and second images based on an attention mechanism, transferring high-frequency textures from the high-resolution image to the low-resolution image. Through block-by-block matching of similar parts, higher precision super-resolution is achieved, generating more realistic texture details. Figure 10 illustrates a schematic diagram of step S210 shown in Figure 2 in one embodiment. As shown in Figure 10, in this embodiment, step S210 may further include steps S1002 to S1008. Step S1002: The first fused image and the first feature-transferred image are fused to obtain a first feature-transferred fused image. In some embodiments, the first fused image and the first feature-transferred image can be fused using multiple residual blocks to obtain the first feature-transferred fused image. Step S1004: The first feature-transferred fused image undergoes super-resolution reconstruction processing to obtain a second reconstructed image. In some embodiments, a spatially-angle-separable convolution method can be used to extract features from the first feature transfer fusion image, followed by upsampling. Then, the image is reconstructed through a convolutional layer and merged with the upsampled first feature transfer fusion image to obtain a second reconstructed image. For example, in Figures 3 and / or 4, the first feature transfer fusion image is used as input to obtain the second reconstructed image. Step S1006: After performing a second downsampling process on the second image, a reverse stretching process is performed based on the second reconstructed image to obtain the second fused image. In some embodiments, the image after the second downsampling process can be reverse stretched based on the second reconstructed image and then merged with the second reconstructed image to generate the second fused image. The image size obtained by the second downsampling process can, for example, be larger than the image size obtained by the first downsampling process. Step S1008: The second fused image and the second feature transfer image are fused to obtain a target multi-view image. For example, in the process shown in Figure 5, replacing 5002 with the first feature transfer fusion image, modifying the downsampling in step S512 to Size×4, and replacing 5006 with the second feature transfer image Size×4, the second feature transfer fusion image can be output. When the network is set to a two-stage cascade, the second feature transfer fusion image is the target multi-view image. According to the cascaded model provided in the embodiments of this disclosure, complementary information from multi-view low-resolution images obtained from light field super-resolution technology and texture features from different resolution images in reference image-based super-resolution technology are fused. It utilizes attention-based feature matching and distortion and transfer between different perspectives from coarse to fine.High-frequency textures can be effectively generated. Multi-scale feature extraction can extract edge features in metal images and generate realistic metal details, solving the problem that the reflective properties of metal images make it difficult to extract edge textures. The method provided in this disclosure achieves large-scale high-precision super-resolution while also having good scalability. Figure 11 illustrates an exemplary three-level cascaded processing procedure. Figure 11 illustrates the overall flowchart of a hybrid multi-resolution cascaded model based on Figures 2 to 10. In Figure 11, the input to the top network is the multi-view LR image 11002, which undergoes three levels of feature extraction, inverse stretching, and fusion of high-frequency features from the corresponding size center-view HR image. The input to the bottom network is the center-view HR image 11004, which extracts depth features of different sizes from the high-resolution center-view image to match the feature map of the low-resolution image, thereby generating three transfer images of different scales, which are then sequentially input into the top network to process the low-resolution image one by one. 15 As shown in Figure 11, the bottom network obtains the first feature map 11006 of the multi-view LR image 11002, the first feature map 11008 (×1) of the center-view HR image 11004, and the second feature map 11008 (×2, ×4, ×8) of the center-view HR image 11004. Based on the first feature map 11006 of the multi-view LR image, the first feature map and the second feature map 11008 of the center-view HR image, feature matching is performed (S1142), followed by upsampling (S1144). Then, based on the similarity 20, the features in the second feature map of the second image are backstretched (S1146) based on the upsampled feature map, and then reconstructed (S1148) into a high-resolution feature transfer image 11010 (Size ×2, Size ×4, Size ×8). After feature extraction of the multi-view LR image 11002 input to the top network in step S1102a, it is sequentially upsampled in step S1105a (S1104), then reconstructed by a convolutional layer and merged with the image after upsampling of the multi-view LR image 11002 (S1103) (S1106). Then, it is downsampled with the center-view LR image 11004 (S1122) (for example, the resulting image size is ×2, which is twice the size of the multi-view LR image 5002 in Figure 5), and then stretched in reverse (S1108) before being merged. Finally, it is fused with the feature transfer image 11010 of size ×2 (S1110a).The first-level feature fusion of the high-resolution center-view image and the low-resolution image is completed. Then, the first-level fusion output is sequentially processed through step S1102b (similar to step S1102a) for feature extraction, followed by step S1105b (similar to step S1105a) for upsampling, reconstruction, and merging. The center-view LR image 11004 is downsampled (S1124) to obtain an image of size ×4 (four times that of the multi-view LR image 5002 in Figure 5) and then inversely stretched. After step S1110b (similar to step S1110a), the feature transfer image 11010 of size ×4 is fused, completing the second-level feature fusion of the high-resolution center-view image and the low-resolution image. The second-stage fusion output is then sequentially processed through step S1102c (similar to step S1102a or step S1102b) for feature extraction. Following step S1105c (similar to step S1105a or step S1105b), it undergoes upsampling, reconstruction, and merging. The center-view LR image 11004 is downsampled (S1126) to obtain a ×8 (eight times the size of the multi-view LR image 5002 in Figure 5) 10 image that is then inversely stretched. This image is then fused with a size ×8 feature transfer image 11010 through step S1110c (similar to step S1110a or step S1110b), resulting in the target super-resolution image 11012. This disclosure proposes a large-scale multi-view metal image super-resolution model based on mixed resolution input. For a set of multi-view images, only low-resolution multi-view images and a single high-resolution center-view image can be acquired, thus balancing shooting costs and technical costs. By alternately extracting spatial and angular features from multi-view images, complementary information from different viewpoints is obtained. Furthermore, by matching the similarity between a high-resolution center viewpoint image and other low-resolution viewpoint images, and transferring high-resolution textures, high-resolution and realistic multi-view images are generated. Simultaneously, complementary images and multi-scale reverse-stretched images are obtained through multi-scale feature extraction and fusion, and then sequentially fused through cascaded connections to generate large-scale high-resolution images. The super-resolution model provided in this disclosure, based on the application of hybrid multi-resolution input to the processing of multi-view metal images, can achieve high-magnification super-resolution. After training with a multi-view metal image training set, this model is more suitable for precision engineering processes. Since cascaded connections and high-frequency feature matching and transfer contribute to achieving high-magnification super-resolution, the model can generate images with higher resolution and more realistic textures when the image magnification is large. Furthermore, due to the use of hybrid resolution input,This method can balance the hardware and software costs of the camera. Figure 12 is a block diagram of a multi-view image processing apparatus according to an exemplary embodiment. The apparatus shown in Figure 12 can be applied, for example, to the server side of the system shown in Figure 1, or to the terminal device of the system shown in Figure 1. Referring to Figure 12, the apparatus 120 provided in this embodiment may include an acquisition module 1202, a super-resolution reconstruction module 1204, a reverse stretching module 1206, a feature matching module 1208, and a fusion module 1210. The acquisition module 1202 can be used to acquire a multi-view image to be processed, which includes a first image and a second image. The first image includes images of at least two views of a target object, and the resolution of the first image is less than the resolution of the second image. The super-resolution reconstruction module 1204 can be used to perform super-resolution reconstruction processing on the first image to obtain a first reconstructed image. The reverse stretching module 1206 can be used to perform a first downsampling process on the second image, and then perform a reverse stretching process based on the first reconstructed image to obtain a first fused image. The first fused image fuses the second image into the first reconstructed image corresponding to at least two viewpoints. The feature matching module 1208 can be used to perform feature matching processing on the first image and the second image to obtain a first feature transfer image. The features of the first feature transfer image are fused with the features of the second image and the features of the first image. The fusion module 1210 can be used to fuse the first fused image and the first feature transfer image to obtain a target multi-view image. The resolution of the target multi-view image is greater than 15 times the resolution of the first image. FIG13 is a block diagram of a multi-view image processing apparatus according to an exemplary embodiment. The apparatus shown in FIG13 can be applied, for example, to the server side of the system shown in FIG1, or to the terminal device of the system shown in FIG1. Referring to Figure 13, the apparatus 130 provided in this embodiment may include an acquisition module 1302, a first feature extraction module 13032, a second feature extraction module 13034, a super-resolution reconstruction module 1304, a reverse stretching module 1306, a feature matching module 1308, and a fusion module 1310. The acquisition module 1302 can be used to acquire a multi-view image to be processed, which includes a first image and a second image. The first image includes images of at least two views of the target object, and the resolution of the first image is lower than the resolution of the second image. The first image may include images of at least three views of the target object, including a central view; the second image may include an image of the central view of the target object. The first feature extraction module 13032 can be used to perform feature extraction processing on the first image.A first feature map of the first image is obtained. The first feature extraction module 13032 can also be used to perform feature extraction processing on the first image using a first encoder (30 HK 30136960 A 20) to obtain the first feature map of the first image. The first encoder includes a convolutional layer and residual blocks. A second feature extraction module 13034 can be used to perform feature extraction processing on the second image to obtain a first feature map and a second feature map of the second image. The size of the second feature map of the second image is larger than the size of the first feature map of the second image. The second feature extraction module 13034 can also be used to perform feature extraction processing on the second image using at least one encoding module of the second encoder to obtain a second feature map of the second image. The second feature map of the second image includes at least two feature maps of different sizes. It can also perform feature extraction processing on the second image using at least two encoding modules of the second encoder to obtain a first feature map of the second image. The encoding modules of the second encoder include convolutional layers and residual blocks. A super-resolution reconstruction module 1304 can be used to perform super-resolution reconstruction processing on the first image to obtain a first reconstructed image. The super-resolution reconstruction module 1304 can also be used to extract features from the first image using a spatially-angle-separable convolution method to obtain a first low-resolution feature map; reconstruct the upsampled first low-resolution feature map through a convolutional layer to obtain a first reconstruction residual map; and merge the first reconstruction residual map with the upsampled first image to obtain a first reconstructed image. The super-resolution reconstruction module 1304 can also be used to perform super-resolution reconstruction processing on the first feature transfer fusion image to obtain a second reconstructed image. The reverse stretching module 1306 can be used to perform reverse stretching processing on the second image based on the first reconstructed image after performing a first downsampling process to obtain a first fused image. The first fused image 20 fuses the second image into the first reconstructed image corresponding to at least two viewpoints. The reverse stretching module 1306 can also be used to perform reverse stretching processing on the second image based on the second reconstructed image after performing a second downsampling process to obtain a second fused image. The feature matching module 1308 can be used to perform feature matching processing on the first image and the second image to obtain a first feature transfer image. The features of the first feature transfer image are fused from the features of the second image 25 and the features of the first image. The feature matching module 1308 can also be used to perform feature matching on the first feature map of the first image and the first feature map of the second image to obtain the similarity between the first feature map of the second image and the first feature map of the first image; based on the similarity, the features in the second feature map of the second image are reverse stretched, and a feature transfer image is reconstructed. The feature transfer image may include the first feature transfer image. The feature transfer image may also include the second feature transfer image.The size of the second feature transfer image is larger than the size of the first feature transfer image. The size of the first feature map of the first image can be the same as the size of the first feature map of the second image. The feature matching module 1308 can also be used to divide the first feature map of the first image into a first number of blocks; divide the first feature map of the second image into a first number of blocks; calculate the similarity between the first feature map of the first image and the first feature map of the second image block by block to obtain a similarity matrix between the first feature map of the first image and the first feature map of the second image. The fusion module 1310 can be used to fuse the first fused image and the first feature transfer image to obtain a target multi-view image, the resolution of which is greater than the resolution of the first image. The fusion module 1310 can also be used to fuse the first fused image and the first feature transfer image to obtain a first feature transfer fused image. The fusion module 1310 can also be used to fuse the second fused image and the second feature transfer image to obtain a target multi-view image. The specific implementation of each module in the device provided in the embodiments of this disclosure can refer to the content of the above method, and will not be repeated here. Figure 14 shows a schematic diagram of the structure of an electronic device in an embodiment of this disclosure. It should be noted that the device shown in FIG14 is only an example of a computer system and should not impose any limitation on the functions and scope of use of the embodiments of this disclosure. As shown in FIG14, the device 1400 includes a central processing unit (CPU) 1401, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1402 or a program loaded from a storage portion 1408 into a random access memory (RAM) 1403. Various programs and data required for the operation of the device 1400 are also stored in the RAM 1403. The CPU 1401, ROM 1402, and RAM 1403 are interconnected via a bus 1404. An input / output (I / O) interface 1405 is also connected to the bus 1404. The following components are connected to I / O interface 1405: an input section 1406 including a keyboard, mouse, etc.; an output section 1407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1408 including a hard disk, etc.; and a communication section 1409 including a network interface card such as a LAN card, a modem, etc. Communication section 1409 performs communication processing via a network such as the Internet. Drive 1410 is also connected to I / O interface 1405 as needed. Removable media 1411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1410 as needed.So that the computer program read therefrom can be installed into storage section 1408 as needed. 5 In particular, according to embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1409, and / or installed from removable medium 1411. When the computer program is executed by central processing unit (CPU) 1401, it performs the functions defined in the system of the present disclosure. It should be noted that the computer-readable medium shown in the present disclosure can be a computer-readable signal medium or a computer-readable storage medium or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof. The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this respect, each block in a flowchart or block diagram may represent a module, program segment, or portion of code.The aforementioned modules, program segments, or portions of code contain one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked 5 in the block may occur in a different order than those shown in the figures. For example, two consecutively indicated blocks may actually be executed substantially in parallel, or sometimes in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram or flowchart, and combinations of blocks in the block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions. The modules described in the embodiments of this disclosure can be implemented in software or hardware. The described modules can also be housed in a processor; for example, a processor may be described as including an acquisition module, a super-resolution reconstruction module, a reverse stretching module, a feature matching module, and a fusion module. The names of these modules do not necessarily constitute a limitation on the module itself; for example, the acquisition module may also be described as "acquiring a multi-view image of mixed resolution to be processed." In another aspect, this disclosure also provides a computer-readable medium, which may be included in the device described in the above embodiments; or it may exist independently and not assembled into the device. The computer-readable medium carries one or more programs that, when executed by the device, cause the device to include: acquiring a multi-view image to be processed, the multi-view image to be processed including a first image and a second image, the first image including images of at least two views of a target object, the resolution of the first image being less than the resolution of the second image; performing super-resolution reconstruction processing on the first image to obtain a first reconstructed image; performing a first downsampling processing on the second image, and then performing a reverse stretching processing based on the first reconstructed image to obtain a first fused image, the first fused image fusing the second image into the first reconstructed image corresponding to the images of at least two views; performing feature matching processing on the first image and the second image to obtain a first feature transfer image, the features of the first feature transfer image fusing features of the second image and features of the first image; fusing the first fused image and the first feature transfer image to obtain a target multi-view image, the resolution of the target multi-view image being greater than the resolution of the first image. This disclosure provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions.This enables the computer device to perform the methods provided in the various optional implementations described above. Exemplary embodiments of this disclosure have been specifically shown and described above. It should be understood that this disclosure is not limited to the detailed structures, arrangements, or implementation methods described herein; rather, this disclosure is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended claims. 10 HK 30136960 A 1 Claim 1. A multi-view image processing method, characterized in that it comprises: acquiring a multi-view image to be processed, the multi-view image to be processed comprising a first image and a second image, the first image comprising images of at least two views of a target object, the resolution of the first image being less than the resolution of the second image; performing super-resolution reconstruction processing on the first image to obtain a first reconstructed image; performing a first downsampling processing on the second image, and then performing a reverse stretching processing based on the first reconstructed image to obtain a first fused image, the first fused image fusing the second image into the first reconstructed image corresponding to the images of the at least two views; performing feature matching processing on the first image and the second image to obtain a first feature transfer image, the features of the first feature transfer image fusing the features of the second image and the features of the first image; fusing the first fused image and the first feature transfer image to obtain a target multi-view image, the resolution of the target multi-view image being greater than the resolution of the first image. 15 2. The method according to claim 1, characterized in that it further comprises: performing feature extraction processing on the first image to obtain a first feature map of the first image; performing feature extraction processing on the second image to obtain a first feature map and a second feature map of the second image, wherein the size of the second feature map of the second image is larger than the size of the 20th feature map of the second image; performing a first feature matching processing on the first image and the second image to obtain a first feature transfer image, comprising: performing feature matching on the first feature map of the first image and the first feature map of the second image to obtain a 25% similarity between the first feature map of the second image and the first feature map of the first image; performing reverse stretching processing on the features in the second feature map of the second image according to the similarity, and reconstructing to obtain a feature transfer image, wherein the feature transfer image includes the first feature transfer image. 3. The method according to claim 2, characterized in that, performing feature matching on the 30th feature map of the first image and the first feature map of the second image to obtain a similarity between the first feature map of the second image and the first feature map of the first image,The method includes: dividing the first feature map of the first image into a first number of blocks; dividing the first feature map of the second image into the first number of blocks; calculating the similarity between the first feature map of the first image and the first feature map of the second image block by block to obtain a similarity matrix between the first feature map of the first image and the first feature map of the second image. 4. The method according to claim 2 or 3, wherein the second feature map of the second image includes at least two feature maps of different sizes; the feature transfer image further includes a second feature transfer image, the size of the second feature transfer image 10 being larger than the size of the first feature transfer image; fusing the first fused image and the first feature transfer image to obtain a target multi-view image, including: fusing the first fused image and the first feature transfer image to obtain a first feature transfer fused image; performing super-resolution reconstruction processing on the first feature transfer fused image to obtain a second reconstructed image; performing a second downsampling processing on the second image, and then performing reverse stretching processing based on the second reconstructed image to obtain a second fused image; fusing the second fused image and the second feature transfer image to obtain a target multi-view image. 5. The method according to claim 2 or 3, characterized in that, performing feature extraction processing on the first image to obtain a first feature map of the first image includes: performing feature extraction processing on the first image through a first encoder to obtain a first feature map of the first image, wherein the first encoder includes a convolutional layer and a residual block; performing feature extraction processing on the second image to obtain a first feature map and a second feature map of the second image includes: performing feature extraction processing on the second image through at least one encoding module of a second encoder to obtain a second feature map of the second image; performing feature extraction processing on the second image through at least two encoding modules of the second encoder to obtain a first feature map of the second image, wherein the encoding module of the second encoder includes a convolutional layer and a residual block. 6. The method according to any one of claims 1 to 3, characterized in that, performing super-resolution reconstruction processing on the first image to obtain a first reconstructed image includes: extracting features from the first image using a spatially-angle-separable convolution method to obtain a first low-resolution feature map; reconstructing the upsampled first low-resolution feature map through a convolutional layer to obtain a first reconstruction residual map; and merging the first reconstruction residual map with the upsampled first image to obtain the first reconstructed image. 7. The method according to any one of claims 1 to 3, characterized in that, the first image includes images of at least three viewpoints of the target object.The at least three perspectives include a central perspective; the second image includes an image of the central perspective of the target object. 8. A multi-view image processing apparatus, characterized in that it comprises: 15 an acquisition module, configured to acquire a multi-view image to be processed, the multi-view image to be processed including a first image and a second image, the first image including images of at least two perspectives of a target object, the resolution of the first image being less than the resolution of the second image; a super-resolution reconstruction module, configured to perform super-resolution reconstruction processing on the first image to obtain a first reconstructed image; 20 a reverse stretching module, configured to perform a first downsampling processing on the second image, and then perform reverse stretching processing based on the first reconstructed image to obtain a first fused image, the first fused image fusing the second image into the first reconstructed image corresponding to the images of the at least two perspectives; 25 a feature matching module, configured to perform feature matching processing on the first image and the second image to obtain a first feature transfer image, the features of the first feature transfer image fusing the features of the second image and the features of the first image; and a fusion module, configured to fuse the first fused image and the first feature transfer image to obtain a target multi-view image, the resolution of the target multi-view image being greater than the resolution of the first image. 30 HK 30136960 A 4 9. The apparatus according to claim 8, characterized in that it further comprises: a first feature extraction module, configured to perform feature extraction processing on the first image to obtain a first feature map of the first image; a second feature extraction module, configured to perform feature extraction processing on the second image to obtain a first feature map and a second feature map of the second image, wherein the size of the second feature map of the second image is larger than the size of the first feature map of the second image; the feature matching module is further configured to perform feature matching on the first feature map of the first image and the first feature map of the second image to obtain a similarity between the first feature map of the second image and the first feature map of the first image; perform reverse stretching processing on the features in the second feature map of the second image according to the similarity, and reconstruct to obtain a feature migration image, wherein the feature migration image includes the first feature migration image. 10. The apparatus according to claim 9, characterized in that the feature matching module is further configured to: divide the first feature map of the first image into a first number of blocks; divide the first feature map of the second image into the first number of blocks; calculate the similarity between the first feature map of the first image and the first feature map of the second image block by block.Obtain the similarity matrix between the first feature map of the first image and the first feature map of the second image. HK 30136960 A 1 Specification Figure 10 106 104 1024 1022 106 104 1024 Figure 1 HK 30136960 A 2 S204 S206 20 S208 S210 S202 Acquire the multi-view image to be processed, which includes a first image and a second image. The first image includes images of at least two views of the target object. The resolution of the first image is less than the resolution of the second image. Perform super-resolution reconstruction processing on the first image to obtain the first reconstructed image. Perform first downsampling processing on the second image, and then perform reverse stretching processing based on the first reconstructed image to obtain a first fused image. The first fused image fuses the second image into the first reconstructed image corresponding to the images of at least two views. Perform feature matching processing on the first image and the second image to obtain a first feature transfer image. The features of the first feature transfer image are fused with the features of the second image and the features of the first image. Fuse the first fused image and the first feature transfer image to obtain the target multi-view image. The resolution of the target multi-view image is greater than the resolution of the first image. Figure 2 S304 S306 S302 uses a spatially and angularly separable convolution method to extract features from the first image, obtaining a first low-resolution feature map. The first low-resolution feature map after upsampling is reconstructed using a convolutional layer, obtaining a first reconstruction residual map. The first reconstruction residual map is then merged with the upsampling first image to obtain the first reconstructed image. Figure 3 HK 30136960 A 3 402 4002. Multi-view LR image 404 406 408 aw*ah am*an am*an aw*ah 4004 Figure 4 S506. Upsampling 5002. Multi-view LR image S512. Downsampling Size ×2 S504. Upsampling S510. Reverse stretching Size ×2 5004. Center-view HR image S502. Feature extraction S508. Merging S514. Fusion 5006. First feature transfer image Figure 5 HK 30136960 A 4 S604 S606 S608 S602 performs feature extraction processing on the first image to obtain a first feature map of the first image. It then performs feature extraction processing on the second image to obtain a first feature map and a second feature map of the second image. The size of the second feature map of the second image is larger than the size of the first feature map of the second image. Feature matching is then performed on the first feature maps of the first and second images to obtain the similarity between them. Based on the similarity, the features in the second feature map of the second image are then subjected to reverse stretching processing.The feature transfer image is reconstructed and obtained. The feature transfer image includes a first feature transfer image (Figure 6). S704. Upsampling S706. Inverse stretching S708. Reconstruction ×1 ×1 ×2 7002. First feature map of multi-view LR image 7004. First feature map of center-view HR image 7006. Second feature map of center-view HR image S702. Feature matching 7008. First feature transfer image (Figure 7). HK 30136960 A 5 S804 S806 S802 The first image is processed by feature extraction from the first encoder to obtain the first feature map of the first image. The first encoder includes a convolutional layer and residual blocks. The second image is processed by feature extraction from at least one encoding module of the second encoder to obtain the second feature map of the second image. The second image is processed by feature extraction from at least two encoding modules of the second encoder to obtain the first feature map of the second image. The encoding module of the second encoder includes a convolutional layer and residual blocks (Figure 8). S906 S904 The first feature map of the second image is divided into a first number of blocks. The similarity between the first feature map of the first image and the first feature map of the second image is calculated block by block to obtain the similarity matrix S902 between the first feature map of the first image and the first feature map of the second image. The first feature map of the first image is divided into a first number of blocks (Figure 9 HK 30136960 A 6 S1004 S1006 S1008 S1002). The first fused image and the first feature-transferred image are fused to obtain the first feature-transferred fused image. The first feature-transferred fused image is subjected to super-resolution reconstruction processing to obtain the second reconstructed image. After the second image is subjected to second downsampling processing, inverse stretching processing is performed based on the second reconstructed image to obtain the second fused image. The second fused image and the second feature-transferred image are fused.To obtain multi-view images of the target (Figure 10), we need to: S1103. Upsampling; S1102. Multi-view LR image; S1103. Output super-resolution image; S1104. Center-view HR image. The process involves: Size ×8, Size ×4, Size ×2. The steps are: Upsampling S1104; Reconstruction S1106; Inverse stretching and fusion S1108; S1110a; S1102a. Feature extraction S1105a; S1102b; Upsampling, reconstruction, inverse stretching and fusion S1105b; S1110b; S1102c; Upsampling, reconstruction, inverse stretching and fusion S1105c; S1110c ×1; 11006 ×1 ×2 ×4 ×8; S1122. Size ×2; S1124. Size ×4; S1126. Size ×8; 11008 S1 1 4 2. Feature matching S1 1 4 4. Upsampling S1 1 4 6. Inverse stretching S1 1 4 8. Reconstruction 11010 Figure 11 HK 30136960 A 7 1202 1204 120 1208 1206 Inverse stretching module Super-resolution reconstruction module Acquisition module Feature matching module 1210 Fusion module Figure 12 1302 1304 130 1308 1306 Inverse stretching module Super-resolution reconstruction module Acquisition module Feature matching module 1310 Fusion module 13032 First feature extraction module 13034 First feature extraction module Figure 13 HK 30136960 A 8 1400 CPU ROM RAM I / O interface Input section Output section Storage section Communication section Driver Removable media 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 Figure 14 HK 30136960 A,
