An image transmission method, device, computer device and storage medium

Abstract

This application relates to an image transmission method, apparatus, computer device, and storage medium. The method includes: performing segmentation processing on a source image to obtain multiple segmented images; generating watermark information data for the multiple segmented images and generating watermark carrier images for the multiple segmented images based on image recognition information of the multiple segmented images and spatiotemporal parameters of the client; performing feature layer extraction on the segmented images and watermark carrier images respectively, generating multiple corresponding segmented feature sub-images and multiple watermark carrier feature sub-images, and fusing them to generate multiple target feature sub-images; determining the upload sequence of the multiple target feature sub-images based on the transmission status information from the client to the server, the watermark information of the target feature sub-images, and the feature level corresponding to the target feature sub-images, and uploading them. This method enables stable transmission of large volumes of watermarked photos in restricted client application environments, thereby ensuring the timeliness and reliability of the watermark information.

Description

Technical Field

[0001] This application relates to the field of image transmission technology, and in particular to an image transmission method, apparatus, computer equipment, and storage medium. Background Technology

[0002] With the development of mobile terminals and network applications, watermarked photos have been widely used in scenarios such as content ownership confirmation, source identification, and business compliance. In existing technologies, the client usually generates a watermarked photo containing time, location, or business identifier after acquiring the photo, and then uploads the watermarked photo to the server for storage or subsequent processing.

[0003] However, in practical applications, the computing power, storage resources, and network environment of client devices often vary significantly. Especially in scenarios with weak networks, network fluctuations, or batch data collection, existing technologies for generating and uploading watermarked photos generally suffer from insufficient stability. On the one hand, existing technologies mostly use the entire watermarked photo or simple fragments as the basic upload object, lacking an adaptive transmission mechanism for the scale of photo data and network conditions. When processing large amounts of photo data, upload delays, failures, or repeated retransmissions are prone to occur, making it difficult to guarantee overall upload efficiency. On the other hand, watermarks are usually generated and fixed in the photo once on the client side. When the upload process is blocked or delayed, the time, location, and other information reflected by the watermark are prone to deviate from the actual business status, thus affecting the timeliness and reliability of the watermark information. Summary of the Invention

[0004] Therefore, it is necessary to provide an image transmission method, device, computer equipment, and storage medium that can achieve stable transmission of large volumes of watermarked photos in a restricted client application environment, thereby ensuring the timeliness and reliability of watermark information.

[0005] On the one hand, an image transmission method is provided, applied to a client, the method comprising: Acquire a source image and perform segmentation processing on the source image to obtain multiple segmented images; Based on the image recognition information of multiple segmented images and the spatiotemporal parameters of the client, watermark information data of multiple segmented images is generated, and watermark carrier images of multiple segmented images are generated based on the watermark information data and preset watermark configuration parameters. Based on multiple preset feature levels, feature layer extraction is performed on the segmented image and the watermark carrier image respectively to generate multiple corresponding segmented feature sub-images and multiple watermark carrier feature sub-images. Based on the relative positions of the multiple segmented feature sub-images in the source image, the watermark mark positions of the multiple watermark carrier feature sub-images are determined, so that the multiple watermark carrier feature sub-images are added to the corresponding positions of the multiple segmented feature sub-images and fused to generate multiple target feature sub-images. Based on the transmission status information from the client to the server, the watermark information of the target feature sub-image, and the feature level corresponding to the target feature sub-image, an upload sequence of multiple target feature sub-images is determined, and multiple target feature sub-images are uploaded to the server based on the upload sequence.

[0006] In one embodiment, the step of acquiring a source image and performing segmentation processing on the source image to obtain multiple segmented images includes: The image size information of the source image is obtained, and preprocessing is performed on the source image based on the image size information to obtain a preprocessed image. The preprocessing includes at least one or more of the following: scaling, cropping, padding, and normalization. Based on the preset segmentation size and the overlap ratio of the segmented images, multiple segmentation regions are determined in the preprocessed image; Based on the multiple segmented regions, multiple corresponding segmented images are cropped from the preprocessed image, and the segmentation identifiers of the multiple segmented images are determined according to the cropping order. Based on the position parameters of the segmented image in the preprocessed image and the segment identifier, the segment association information of the segmented image is generated.

[0007] In one embodiment, generating watermark information data for multiple sliced ​​images based on image recognition information of multiple sliced ​​images and spatiotemporal parameters of the client includes: Based on a preset image recognition model, image recognition processing is performed on multiple sliced ​​images to obtain image recognition information of multiple sliced ​​images. The image recognition information includes at least one or more of the following: object category, scene label, text recognition result, and object position parameters. Obtain the spatiotemporal parameters of the client, at least the time parameter and the geographical location parameter; Determine the categories of hot business objects and the boundary information of hot areas based on preset business association parameters; Based on the geographic location parameters and the object location parameters, the location information of the object in the segmented image is fused and determined to obtain object location association information; The object location association information is compared with the hotspot region boundary information to determine whether the object in the segmented image is located in the preset hotspot region, and hotspot region discrimination information is generated. The object category is compared with the business hotspot object category to determine whether the object in the segmented image is a business hotspot object, and hotspot object discrimination information is generated; One or more of the image recognition information, spatiotemporal parameters, hotspot region discrimination information, and hotspot object discrimination information of the segmented image are combined to determine the watermark payload of the segmented image, and the watermark payload is encoded to generate the watermark information data of the segmented image.

[0008] In one embodiment, the step of performing feature layer extraction on the segmented image and the watermark carrier image based on multiple preset feature levels to generate corresponding multiple segmented feature sub-images and multiple watermark carrier feature sub-images includes: A first feature extraction network is constructed based on multiple preset feature levels, and the segmented image is input into the first feature extraction network; Based on the feature extraction path in the first feature extraction network, multiple convolution operations are performed on the segmented image, and after at least one convolution operation, downsampling operation is performed on the feature map obtained by the convolution operation to generate multiple segmented feature sub-maps corresponding to the segmented image, and the multiple segmented feature sub-maps correspond one-to-one with the multiple feature levels; A second feature extraction network is constructed based on multiple feature levels consistent with the first feature extraction network; The watermark carrier image is subjected to frequency domain transformation to obtain a frequency domain coefficient map, and the frequency domain coefficient map is input into the second feature extraction network; Based on the feature extraction path in the second feature extraction network, a convolution operation is performed on the frequency domain coefficient map, and pooling and / or downsampling operations are performed on the feature map obtained by the convolution operation to generate multiple watermark intermediate images corresponding to the watermark carrier image. The multiple watermark intermediate images correspond to multiple feature levels. Scale alignment processing is performed on multiple watermark intermediate images to generate multiple watermark carrier feature sub-images corresponding to the watermark carrier image, so that the multiple segment feature sub-images corresponding to the segmented images correspond one-to-one with the multiple watermark carrier feature sub-images corresponding to the watermark carrier image at multiple feature levels.

[0009] In one embodiment, determining the watermark marker positions of the plurality of watermark carrier feature sub-images based on their relative positions in the source image includes: Obtain the segmentation position information of multiple segmented images in the source image, and determine the relative positional relationship of multiple segmented feature sub-images in the source image based on the segmentation position information; Based on the relative positional relationship, the spatial mapping relationship between the segmented feature sub-image and the source image is hierarchically aligned at multiple feature levels to generate hierarchical positional mapping results corresponding to multiple feature levels; Based on the hierarchical position mapping result, the candidate watermark mark positions of multiple watermark carrier feature sub-images on the source image are determined, so that the relative position relationship between the candidate watermark mark positions of multiple watermark carrier feature sub-images and multiple segmented feature sub-images remains consistent. Boundary constraints and / or conflict resolution processes are performed on multiple candidate watermark locations to generate multiple watermark location sub-maps of the watermark carrier.

[0010] In one embodiment, determining an upload sequence for multiple target feature sub-images based on the client-to-server transmission status information, the watermark information of the target feature sub-image, and the feature level corresponding to the target feature sub-image, and uploading the multiple target feature sub-images to the server based on the upload sequence, includes: Obtain and parse the transmission status information from the client to the server to obtain transmission association parameters; A transmission status index is generated based on the transmission association parameters to at least characterize the transmission status level from the client to the server; The watermark information of the target feature sub-image is obtained and parsed to obtain the watermark payload elements; Semantic recognition is performed on the watermark payload elements to generate semantic recognition results, and watermark semantic indicators are determined based on the correlation between the semantic recognition results and preset key semantics. Based on the feature hierarchy of the target feature sub-image, a mapping is performed to determine the feature hierarchy index of the target feature sub-image, wherein the feature hierarchy index is positively correlated with the abstract scale level of the target feature sub-image; A first weight is obtained by performing a positive mapping based on the transmission status index, so as to at least characterize the contribution of the watermark semantic index to the upload priority. A second weight is obtained by performing a reverse mapping based on the transmission status index, so as to at least characterize the contribution of the feature level index to the upload priority. Based on the first weight, the second weight, the watermark semantic index, and the feature level index, the upload priority index of the target feature sub-image is determined; The target feature sub-graphs are sorted according to the multiple upload priority indicators to generate an initial upload sequence; Based on preset sequence local optimization rules, the initial upload sequence is subjected to sequence local optimization processing to generate a target upload sequence, and multiple target feature sub-images are uploaded to the server based on the target upload sequence; The sequence local optimization process includes at least one or more of the following: performing adjacent aggregation and rearrangement on multiple target feature sub-images corresponding to the same slice image so that the target feature sub-images of the same slice image at multiple feature levels form a continuous upload segment; and performing regional continuity constraint rearrangement based on slice location information so that the target upload sequence preferentially covers the target feature sub-images corresponding to spatially adjacent slice images.

[0011] On the other hand, an image transmission method is provided, applied to a server, the method comprising: Receive multiple target feature sub-images uploaded by the client, and determine the feature level corresponding to the multiple target feature sub-images and the segment identifier of the corresponding segmented image based on the description information associated with the target feature sub-images; Based on the feature hierarchy, multiple target feature sub-images are aggregated, and according to the spatial relative position relationship indicated by the segmentation identifier, the aggregated multiple target feature sub-images are spliced ​​to generate intermediate feature images corresponding to multiple feature levels; Based on preset integrity conditions, image integrity verification is performed on multiple intermediate feature maps; In response to one or more of the intermediate feature maps failing the image integrity check, the intermediate feature maps are compared across multiple feature levels to generate a stitching comparison result. Based on the stitching comparison result, missing inference and fragment reconstruction processing are performed on the intermediate feature maps that failed the check to fill in the missing regions and update the intermediate feature maps, until all of the intermediate feature maps pass the image integrity check. Perform cross-feature-level fusion processing on multiple intermediate feature maps that have passed the verification to generate a fused feature map, and generate a target image based on the fused feature map.

[0012] In another aspect, an image transmission apparatus is characterized in that the apparatus includes a client and a server for implementing any of the above-described image transmission methods, the server comprising: The receiving module is used to receive multiple target feature sub-images uploaded by the client, and determine the feature level corresponding to the multiple target feature sub-images and the segment identifier of the corresponding segmented image based on the description information associated with the target feature sub-images. The stitching module is used to collect multiple target feature sub-images based on the feature level, and perform stitching processing on the collected multiple target feature sub-images according to the spatial relative position relationship indicated by the segmentation identifier, to generate intermediate feature images corresponding to multiple feature levels; The verification module is used to perform image integrity verification on multiple intermediate feature maps based on preset integrity conditions; The completion module is configured to, in response to one or more of the intermediate feature maps failing the image integrity check, compare the multiple intermediate feature maps across multiple feature levels, generate a stitching comparison result, and perform missing inference and fragment reconstruction processing on the intermediate feature maps that failed the check based on the stitching comparison result, so as to complete the missing regions and update the intermediate feature maps, until all the intermediate feature maps pass the image integrity check. The generation module is used to perform cross-feature level fusion processing on multiple intermediate feature maps that have passed the verification, generate a fused feature map, and generate a target image based on the fused feature map.

[0013] In another aspect, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the following steps: Acquire a source image and perform segmentation processing on the source image to obtain multiple segmented images; Based on the image recognition information of multiple segmented images and the spatiotemporal parameters of the client, watermark information data of multiple segmented images is generated, and watermark carrier images of multiple segmented images are generated based on the watermark information data and preset watermark configuration parameters. Based on multiple preset feature levels, feature layer extraction is performed on the segmented image and the watermark carrier image respectively to generate multiple corresponding segmented feature sub-images and multiple watermark carrier feature sub-images. Based on the relative positions of the multiple segmented feature sub-images in the source image, the watermark mark positions of the multiple watermark carrier feature sub-images are determined, so that the multiple watermark carrier feature sub-images are added to the corresponding positions of the multiple segmented feature sub-images and fused to generate multiple target feature sub-images; Based on the transmission status information from the client to the server, the watermark information of the target feature sub-image, and the feature level corresponding to the target feature sub-image, an upload sequence of multiple target feature sub-images is determined, and multiple target feature sub-images are uploaded to the server based on the upload sequence.

[0014] In another aspect, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, performs the following steps: Acquire a source image and perform segmentation processing on the source image to obtain multiple segmented images; Based on the image recognition information of multiple segmented images and the spatiotemporal parameters of the client, watermark information data of multiple segmented images is generated, and watermark carrier images of multiple segmented images are generated based on the watermark information data and preset watermark configuration parameters. Based on multiple preset feature levels, feature layer extraction is performed on the segmented image and the watermark carrier image respectively to generate multiple corresponding segmented feature sub-images and multiple watermark carrier feature sub-images. Based on the relative positions of the multiple segmented feature sub-images in the source image, the watermark mark positions of the multiple watermark carrier feature sub-images are determined, so that the multiple watermark carrier feature sub-images are added to the corresponding positions of the multiple segmented feature sub-images and fused to generate multiple target feature sub-images; Based on the transmission status information from the client to the server, the watermark information of the target feature sub-image, and the feature level corresponding to the target feature sub-image, an upload sequence of multiple target feature sub-images is determined, and multiple target feature sub-images are uploaded to the server based on the upload sequence.

[0015] The aforementioned image transmission method, apparatus, computer equipment, and storage medium, by performing segmentation processing on the source image on the client side and generating corresponding watermark information data and watermark carrier images based on each segmented image, organize and embed the watermark information at the segmented granularity, thereby avoiding the problem of excessive data volume caused by uploading the entire watermarked photo as a single object. Furthermore, feature extraction is performed on the segmented images and watermark carrier images at multiple preset feature levels, and the corresponding fusion of watermark carrier feature sub-images and segmented feature sub-images is completed at the feature levels. This allows the watermark information to be combined with the image content in the form of feature sub-images, thereby improving the image quality while ensuring the relevance of the watermark information. The system enhances the flexibility and controllability of image data transmission. Furthermore, during the upload phase, it comprehensively considers the transmission status information from the client to the server, the watermark information of the target feature sub-image, and the corresponding feature level to dynamically determine the upload sequence of the target feature sub-image. This allows for the priority upload of target feature sub-images with higher business value or higher abstraction levels, even under unstable network conditions or bandwidth limitations. This reduces the overall risk of transmission failure and improves the effective transmission rate of image data. Consequently, even in weak network or batch upload scenarios, it can gradually provide the server with effective image features containing watermark information, minimizing the adverse effects of upload delays or interruptions on the timeliness and reliability of the watermark. Attached Figure Description

[0016] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is an application environment diagram of the image transmission method in one embodiment; Figure 2 This is a flowchart illustrating an image transmission method applied to a client in one embodiment; Figure 3 This is a flowchart illustrating an image transmission method applied to a server in one embodiment; Figure 4 This is a structural block diagram of an image transmission device in one embodiment; Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] The image transmission method provided in this application can be applied to, for example... Figure 1 In the application environment shown, client 102 communicates with server 104 via a network. Client 102 is used to acquire source images and process them according to the method of this application, including but not limited to: segmenting the source image, generating watermark information data corresponding to the segmented images, performing feature-level processing on the segmented images and the watermark carrier image based on the watermark information, and forming a target feature sub-image for uploading; at the same time, client 102 is also used to control the uploading order of the target feature sub-image according to the transmission status information from the client to the server, and upload the target feature sub-image to server 104; wherein, client 102 may be, but is not limited to, a personal computer, laptop, smartphone, tablet, portable wearable device or other terminal device with image processing and network communication capabilities, and is not limited to specific hardware form or operating system; Server 104 receives multiple target feature sub-images uploaded by client 102, and collects, stitches, and verifies the target feature sub-images based on the feature hierarchy information and segmentation identifiers associated with them. If the target feature sub-images are incomplete, server 104 can also perform missing inference and incomplete image reconstruction based on the relationships between different feature levels to generate the target image. Server 104 can be implemented by a standalone server or by a server cluster consisting of multiple servers to adapt to different concurrency scales and business processing needs. This application does not limit the specific deployment method of server 104.

[0020] In one embodiment, such as Figure 2 As shown, an image transmission method is provided, which is applied to Figure 1 Client 102 in the middle includes: Step 201: Obtain the source image and perform segmentation processing on the source image to obtain multiple segmented images; Step 202: Based on the image recognition information of multiple segmented images and the spatiotemporal parameters of the client, generate watermark information data of multiple segmented images, and generate watermark carrier images of multiple segmented images based on the watermark information data and preset watermark configuration parameters. Step 203: Based on multiple preset feature levels, perform feature layer extraction on the sliced ​​image and the watermark carrier image respectively to generate multiple sliced ​​feature sub-images and multiple watermark carrier feature sub-images. Step 204: Determine the watermark mark position of multiple watermark carrier feature sub-images based on the relative positions of multiple segmented feature sub-images in the source image, so as to add multiple watermark carrier feature sub-images to the corresponding positions of multiple segmented feature sub-images and fuse them to generate multiple target feature sub-images. Step 205: Based on the transmission status information from the client to the server, the watermark information of the target feature sub-image, and the feature level corresponding to the target feature sub-image, determine the upload sequence of multiple target feature sub-images, and upload the multiple target feature sub-images to the server based on the upload sequence.

[0021] In the aforementioned image transmission method, the source image is segmented on the client side, and corresponding watermark information data and watermark carrier images are generated based on each segment image. This allows the watermark information to be organized and embedded at the segment granularity, thus avoiding the problem of excessive data volume caused by uploading the entire watermarked photo as a single object. Furthermore, feature extraction is performed on the segmented images and watermark carrier images at multiple preset feature levels, and the corresponding fusion of watermark carrier feature sub-images and segment feature sub-images is completed at the feature levels. This allows the watermark information to be combined with the image content in the form of feature sub-images, thereby improving the image data transmission efficiency while ensuring the correlation of watermark information. The system enhances flexibility and controllability during the transmission process. Furthermore, by comprehensively considering the transmission status information from the client to the server, the watermark information of the target feature sub-image, and the corresponding feature level during the upload phase, the upload sequence of the target feature sub-image is dynamically determined. This allows for the priority upload of target feature sub-images with higher business value or higher abstraction levels even under unstable network conditions or limited bandwidth, thereby reducing the overall transmission failure risk and improving the effective transmission rate of image data. Consequently, even in weak network or batch upload scenarios, the system can gradually provide the server with effective image features containing watermark information, reducing the adverse effects of upload delays or interruptions on the timeliness and reliability of the watermark.

[0022] In one embodiment, a source image is acquired and sliced ​​to obtain multiple sliced ​​images, including: Obtain the image size information of the source image, and perform preprocessing on the source image based on the image size information to obtain a preprocessed image. The preprocessing includes at least one or more of the following: scaling, cropping, padding, and normalization. The image size information includes at least the image width and height, and is preferably obtained by reading the image file header information, decoder output parameters or image metadata; normalization refers to the process of mapping pixel values ​​to a target numerical range according to a preset range, and preferably the pixel values ​​are mapped to [0,1] or [-1,1]; Based on the preset segmentation size and the overlap ratio of the segmented images, multiple segmentation regions are determined in the preprocessed image; Wherein, the segment size refers to the width and height parameters of the segmented region. Preferably, the segment size can be set to 256×256, 512×512, or 1024×1024 pixels. The segment image overlap ratio refers to the ratio of the overlap width between adjacent segmented regions to the segment size. Preferably, the overlap ratio can be set to 0% to 30%, for example, 10% or 20%. Specifically, the segmented region can be obtained in the following way: the step size is obtained according to step size = segment size × (1 - overlap ratio), and the window is slid in rows and columns on the preprocessed image according to the step size. For regions whose edges are insufficient to fill the entire window, preferably, edge filling can be performed on the preprocessed image or alignment and cropping can be performed on the last window. Based on multiple segmented regions, multiple corresponding segmented images are cropped from the preprocessed image, and the segmentation identifiers of the multiple segmented images are determined according to the cropping order. The segment identifier refers to an identifier used to uniquely identify a segmented image; preferably, the segment identifier can be obtained through row and column index encoding. Based on the position parameters of the segmented images in the preprocessed image and the segment identifier, the segment association information of the segmented images is generated; The location parameters include at least the coordinates of the top left corner of the segmented region and the width and height of the segmented region; the segment association information refers to the structured description information used for server-side reconstruction and positioning, including at least one or more of the following: source image identifier, segment identifier, location parameters, preprocessing scaling ratio parameters, and overlap ratio parameters; preferably, the segment association information can be encapsulated in TLV structure, JSON structure or binary structure.

[0023] Specifically, in this embodiment, after acquiring the source image, image size-based preprocessing is introduced, and the image is segmented by combining a preset segment size and segment overlap ratio. At the same time, segment association information containing location information and segment identifier is generated for each segment image, so that the segment images have clear global positioning relationships while maintaining spatial continuity. This not only avoids the problem of spatial information loss or difficulty in reconstruction caused by simple segmentation in the prior art, but also provides a reliable basis for the independent transmission, sequential control and server reconstruction of subsequent segment images, thereby improving the controllability of transmission and reconstruction of large-size images in complex network environments from an engineering perspective.

[0024] In one embodiment, watermark information data for multiple image segments is generated based on image recognition information of multiple image segments and spatiotemporal parameters of the client, including: Based on a preset image recognition model, image recognition processing is performed on multiple slice images to obtain image recognition information for multiple slice images. The image recognition information includes at least one or more of the following: object category, scene label, text recognition result, and object location parameters. Here, the image recognition model refers to the model used to output the semantics of objects and scenes. Preferably, the image recognition model can be a combination of a convolutional neural network model, a visual Transformer model, or an OCR model; the object position parameter is preferably a bounding box parameter. Obtain the client's spatiotemporal parameters, at least the time parameter and the geographic location parameter; The time parameter is preferably a UTC timestamp or a local timestamp; preferably, the time parameter can be obtained through a system clock or network time synchronization; the geographic location parameter includes at least longitude and latitude, and may optionally include altitude, positioning accuracy and positioning method; preferably, the geographic location parameter can be obtained through GNSS / GPS, base station positioning, WiFi positioning or IP coarse positioning. Determine the categories of hot business objects and the boundary information of hot areas based on preset business association parameters; Among them, the preset business association parameters refer to the association rules and data set configured on the business side; preferably, the business hotspot object category can be a preset category list (such as "vehicles / stores / billboards / certificates") or a category ID set; the hotspot area boundary information refers to the spatial range description of the hotspot area, preferably a polygonal boundary, a circular boundary (center point + radius) or a grid boundary; preferably, the business association parameters can be issued by the server and cached on the client, or loaded by the local configuration file; The location information of objects in the segmented image is determined by fusing geographic location parameters and object location parameters to obtain object location association information; Among them, the object location association information refers to the structured record that establishes a relationship between "the position of the object in the image" and "the geographical location / time of the acquisition"; preferably, the object location association information includes at least: object category, object location parameters, geographical location parameters, and time parameters; optionally, it includes device orientation, focal length, or shooting posture parameters; The object location association information is compared with the hot spot region boundary information to determine whether the object in the segmented image is located in the preset hot spot region, and hot spot region discrimination information is generated. Specifically, the comparison can be obtained in the following ways: when the hot spot area is a polygon boundary, the point is determined to be inside the polygon for the geographic location parameter; when the hot spot area is a circular boundary, the distance between the geographic location and the center point is calculated and compared with the radius threshold; preferably, the distance threshold can be set to 50m, 100m or 200m. The object category is compared with the business hotspot object category to determine whether the object in the segmented image is a business hotspot object, and hotspot object discrimination information is generated. The hotspot object discrimination information is preferably a binary label or confidence score. One or more of the image recognition information, spatiotemporal parameters, hotspot region discrimination information, and hotspot object discrimination information of the segmented image are combined to determine the watermark payload of the segmented image, and the watermark payload is encoded to generate watermark information data of the segmented image. Wherein, the watermark payload refers to the set of effective payload fields to be embedded or associated; preferably, the watermark payload includes at least one of the following: fragment identifier, time parameter, geographic location parameter, object category / scene label and hotspot discrimination result; preferably, the encoding process includes one or more of the following: field serialization (TLV / CBOR / JSON), error correction encoding (e.g., RS code), encryption or signature (e.g., HMAC), and verification digest (e.g., CRC).

[0025] Specifically, in this embodiment, watermark information data is generated separately for each image segment, and the image recognition information is combined and encoded with the client's spatiotemporal parameters and business hotspot discrimination results. This makes the watermark payload correspond to the specific segment content and its business semantics, so that the watermark information no longer depends on the unified generation of the whole image, but can exist independently with the image segment and participate in the transmission. Thus, even in the scenario of asynchronous or delayed upload of segments, the consistency between the watermark information and the actual collected content and business status can still be guaranteed, and the problem of watermark timeliness deterioration caused by overall upload delay can be alleviated.

[0026] In one embodiment, feature extraction is performed on the segmented image and the watermark carrier image based on multiple preset feature levels to generate multiple corresponding segmented feature sub-images and multiple watermark carrier feature sub-images, including: A first feature extraction network is constructed based on multiple preset feature levels, and the sliced ​​image is input into the first feature extraction network; Among them, the feature level refers to the feature layer with different spatial scales / semantic abstraction levels. Preferably, the feature level can correspond to different downsampling ratios, such as 1 / 4, 1 / 8, 1 / 16 or 1 / 32. Preferably, the first feature extraction network can be a convolutional network or a Transformer network that includes a backbone network and multi-scale output heads. Based on the feature extraction path in the first feature extraction network, multiple convolution operations are performed on the segmented image, and after at least one convolution operation, downsampling operation is performed on the feature map obtained by the convolution operation to generate multiple segmented feature sub-maps corresponding to the segmented image. The multiple segmented feature sub-maps correspond one-to-one with multiple feature levels. Among them, the downsampling operation is preferably stride convolution, pooling, or interpolation downsampling; A second feature extraction network is constructed based on multiple feature levels consistent with the first feature extraction network. The number and scale of the output levels of the second feature extraction network are consistent with those of the first feature extraction network to ensure alignment of subsequent levels. The watermark carrier image is subjected to frequency domain transformation to obtain a frequency domain coefficient map, and the frequency domain coefficient map is input into the second feature extraction network. Among them, the watermark carrier image refers to the image / pattern carrying the watermark code or watermark template, preferably a preset template image, an identifier image or a random texture image; the frequency domain transformation processing preferably includes DCT, DFT or wavelet transform, and the frequency domain coefficient image can be an amplitude spectrum / phase spectrum or a block coefficient matrix. Based on the feature extraction path in the second feature extraction network, convolution operation is performed on the frequency domain coefficient map, and pooling and / or downsampling operation is performed on the feature map obtained by the convolution operation to generate multiple intermediate watermark images corresponding to the watermark carrier image. The multiple intermediate watermark images correspond to multiple feature levels; whereby, the intermediate watermark image refers to the intermediate representation of the frequency domain features at each scale. Scale alignment processing is performed on multiple intermediate watermark images to generate multiple watermark carrier feature sub-images corresponding to the watermark carrier image, so that the multiple segment feature sub-images corresponding to the segmented image correspond one-to-one with the multiple watermark carrier feature sub-images corresponding to the watermark carrier image at multiple feature levels. The scale alignment process preferably includes upsampling / downsampling and channel number alignment (e.g., 1×1 convolution); preferably, the spatial size error after scale alignment does not exceed 1 pixel or does not exceed a preset tolerance threshold.

[0027] Specifically, in this embodiment, feature extraction is performed on the sliced ​​image and the watermark carrier image at multiple feature levels, and sliced ​​feature sub-images and watermark carrier feature sub-images that correspond to each other in scale and level are generated. This allows image information to form independently processable data units at different abstraction levels. As a result, the processing process that originally depended on the whole image can be transformed into a multi-level, separable feature processing process, providing a structural foundation for subsequent selective transmission and phased reconstruction according to feature levels. This avoids the problem in the prior art where whole image processing is difficult to progressively recover in a weak network environment.

[0028] In one embodiment, determining the watermark marker positions of multiple watermark carrier feature sub-images based on their relative positions in the source image includes: Obtain the segmentation position information of multiple segmented images in the source image, and determine the relative positional relationship of multiple segmented feature sub-images in the source image based on the segmentation position information; The segment location information preferably comes from the segment association information, including at least the starting position and coverage area of ​​the segment region in the source image; the relative position relationship refers to the spatial adjacency or overlap relationship between segments, preferably including one or more of the following: vertical adjacency, horizontal adjacency, diagonal adjacency, and overlapping coverage relationship; Based on the relative positional relationship, the spatial mapping relationship between the segmented feature sub-image and the source image is hierarchically aligned at multiple feature levels to generate hierarchical positional mapping results corresponding to multiple feature levels; The hierarchical position mapping result refers to the mapping result of converting the "segment position range in the coordinate domain of the source image" into the "position range in the coordinate domain of the corresponding feature map of the feature level". Preferably, the hierarchical alignment can be obtained in the following way: first, determine the resolution reduction relationship of the feature level relative to the source image (i.e., how many times the feature map of the level is reduced in the width and height directions relative to the source image), then convert the starting position and coverage size of the segmented region proportionally according to the reduction relationship, and perform discretization processing on the conversion result to align it to the grid index of the feature map, thereby obtaining the corresponding mapping position range under the level; wherein, the discretization processing is preferably rounding down or rounding to the nearest integer to ensure that the mapping range falls on the valid index; Based on the hierarchical location mapping results, the candidate watermark marker positions of multiple watermark carrier feature sub-images on the source image are determined so that the candidate watermark marker positions of multiple watermark carrier feature sub-images are consistent with the relative positional relationship of multiple segment feature sub-images. Among them, the candidate watermark mark position refers to the region or center position on the corresponding layer feature map used to place the watermark mark, preferably the center region or key texture region of the segmented mapping range. Perform boundary constraints and / or conflict resolution on multiple candidate watermark locations to generate watermark locations for multiple watermark carrier feature sub-maps. Boundary constraints refer to the constraints that limit the marker position to be within the effective area and maintain a minimum safe distance from the edge. Preferably, the minimum safe distance can be set to a number of grid cells inward from the feature map edge, such as one to eight yards. Conflict resolution refers to the adjustment strategy when multiple candidate positions overlap or are too close together. Preferably, it includes: retaining the marker position with higher priority and performing neighborhood translation on the remaining positions; or performing a suppression strategy in the candidate set so that those with higher overlap are replaced by nearby available positions. The degree of overlap is preferably measured by the "ratio of overlapping coverage area to merged coverage area". For example, the overlap ratio threshold can be set to about 30%, and the spacing threshold can be set to a number of feature grid cells.

[0029] Specifically, in this embodiment, by utilizing the relative position information of the segmented images in the source image, the position mapping of the segmented feature sub-images and the watermark carrier feature sub-images is performed at multiple feature levels, and the watermark mark position is determined on this basis. This ensures that the watermark information maintains a consistent spatial correspondence at different feature levels, avoiding the problem of the watermark embedding position being disconnected from the image structure in the prior art. This allows for reasonable reconstruction based on the spatial mapping relationship even if some feature sub-images are missing or delayed, thereby improving the positional consistency during multi-level feature fusion.

[0030] In one embodiment, based on the client-to-server transmission status information, the watermark information of the target feature sub-image, and the feature level corresponding to the target feature sub-image, an upload sequence of multiple target feature sub-images is determined, and the multiple target feature sub-images are uploaded to the server based on the upload sequence, including: Obtain and parse the transmission status information from the client to the server to obtain the transmission-related parameters; Among them, transmission status information refers to status data reflecting link quality, which can preferably be obtained from network protocol stack statistics, application layer detection, or server-side receipt summaries; transmission-related parameters include at least one or more of throughput, round-trip time, packet loss level, retransmission level, and jitter level. A transmission status index is generated based on transmission correlation parameters to at least characterize the transmission status level from the client to the server. The transmission status index represents a comprehensive evaluation of the current link status, preferably a discrete level or a continuous score. Specifically, it can be obtained as follows: first, the throughput capacity is converted into a standardized score of "the larger the better," then the round-trip delay, packet loss level, and jitter level are converted into standardized scores of "the smaller the better," and then weighted and synthesized according to preset weights to obtain a comprehensive score. Preferably, the comprehensive score can be divided into three or five levels, for example, into three levels: "good," "average," and "poor." The watermark information of the target feature sub-graph is obtained and parsed to obtain the watermark payload elements; wherein, the watermark payload elements refer to a set of fields that can be used for semantic discrimination and business association, including at least one or more of the following: object category, scene label, hotspot discrimination result, time stamp, and location stamp; Semantic recognition is performed on the watermark payload elements to generate semantic recognition results. Based on the correlation between the semantic recognition results and preset key semantics, the watermark semantic index is determined. The preset key semantics refer to the set of semantics of concern configured on the business side, preferably issued by the server and updated periodically, such as a list of hot object categories, a set of key scene tags, a library of key text keywords, or key area tags. The logic for calculating the correlation can be described as follows: when the recognition result hits the set of key semantics, a higher correlation is output; otherwise, a lower correlation is output. If the recognition model provides confidence, the confidence and the hit result can be combined to obtain the watermark semantic index. Mapping is performed based on the feature hierarchy of the target feature sub-image to determine the feature hierarchy index of the target feature sub-image. The feature hierarchy index is positively correlated with the abstract scale level of the target feature sub-image. The feature hierarchy index refers to the abstract scale level of the target feature sub-image, preferably represented by a hierarchy number or hierarchy category. Preferably, the hierarchy that is more conducive to forming a quick preview and basic structure can be set to a higher hierarchy importance. A forward mapping is performed based on the transmission status indicators to obtain the first weight, which at least represents the contribution of the watermark semantic indicators to the upload priority; whereby forward mapping means that the better the link, the more likely it is to be uploaded according to semantic importance; preferably, the first weight can be larger in the "good" level and smaller in the "poor" level. A second weight is obtained by performing a reverse mapping based on the transmission status indicators, which at least represents the contribution of the feature level indicators to the upload priority. The reverse mapping means that the worse the link, the more likely it is to be uploaded first, which helps the server to quickly form the intermediate feature map connectivity coverage of the basic level data. Preferably, the second weight can be larger in the "poor" level and smaller in the "good" level. Based on the first weight, the second weight, the watermark semantic index, and the feature hierarchy index, the upload priority index of the target feature sub-image is determined. Specifically, the calculation logic of the upload priority index can be described as follows: calculate the "semantic contribution score" and the "hierarchical contribution score" respectively, where the semantic contribution score is determined by the first weight and the watermark semantic index, and the hierarchical contribution score is determined by the second weight and the hierarchical importance. Then, the two types of contribution scores are combined to obtain the final priority index. Multiple target feature sub-graphs are sorted based on multiple upload priority indicators to generate an initial upload sequence; Based on preset sequence local optimization rules, the initial uploaded sequence is subjected to sequence local optimization processing to generate the target uploaded sequence, and multiple target feature sub-graphs are uploaded to the server based on the target uploaded sequence. The sequence local optimization rules are used to improve the continuity and spatial continuity of the same piece without significantly changing the global priority.

[0031] Specifically, in this embodiment, by combining the transmission status from the client to the server, the watermark semantics of the target feature sub-image, and the feature level information, the upload priority of each target feature sub-image is dynamically determined, and local optimization processing is performed on the initial upload sequence, so that the upload process can be adaptively adjusted under different network conditions. Thus, when transmission conditions are limited, target feature sub-images with higher semantic importance or lower abstraction level can be uploaded first, ensuring that the server obtains data that plays a key role in the recovery of the overall image and watermark information as early as possible, avoiding the problem of long-term lag of key information due to indiscriminate uploading in the prior art.

[0032] In a preferred embodiment, based on preset sequence local optimization rules, sequence local optimization processing is performed on the initial upload sequence to generate a target upload sequence, including: An initial upload sequence is obtained, and scheduling entries are generated for multiple watermark feature sub-graphs in the initial upload sequence. Each scheduling entry refers to a structured record used for sequence scheduling and includes at least one or more of the following: fragment identifier, feature level identifier, fragment spatial coordinate parameters, data volume parameters, and estimated transmission delay parameters. Preferably, the data volume description can be obtained by counting the byte length after encoding the feature sub-graphs; the estimated transmission delay description can be obtained by superimposing the round-trip delay estimate on the ratio of the data volume to the available bandwidth estimate. Based on a preset window length, a sliding segmentation process is performed on the initial uploaded sequence to obtain at least one local sequence segment. An adjacency table is constructed for each local sequence segment. The preset window length refers to the number of entries participating in each local reordering, preferably set to eight to sixty-four, such as sixteen or thirty-two. An adjacency table is constructed for each local sequence segment. The adjacency table refers to a set of records used to characterize the switching relationship between adjacent scheduling entries within a local sequence segment, including at least whether the segment is switched, whether the level is switched, and the degree of spatial location difference. For multiple local sequence segments, a local evaluation value is calculated. The local evaluation value refers to a set of computable indicators used to measure the quality of a local sequence segment. Specifically, the calculation of the local evaluation value includes at least the following: counting the number of times the segment identifier changes in adjacent entries as the segment switching cost; counting the number of times the feature level changes in adjacent entries as the level switching cost; accumulating the spatial jump cost based on the degree of change in the spatial distance between the corresponding segments of adjacent entries in the source image; and estimating the preview gain based on the degree of new coverage of the basic level or key region in this local sequence. Under the premise of satisfying the preset perturbation constraints, local transformation processing is performed on multiple local sequence segments to generate candidate local sequence segments. The preset perturbation constraints refer to the conditions that limit the magnitude of local rearrangement. Preferably, the movement distance of any entry relative to its original position is limited to a certain number of positions, such as no more than 3-5 positions. The local transformation processing includes at least adjacent swapping, block insertion, and group convergence. Group convergence refers to moving entries with the same fragment identifier into continuous sub-segments as much as possible, and performing progressive rearrangement within the continuous sub-segments according to the feature level from basic to high level. The local evaluation value of the candidate local sequence segments is recalculated, and the local sequence segments are updated according to the preset selection criteria. The preset selection criteria include at least: reducing the cost of segment switching and / or reducing the cost of hierarchical switching and / or reducing the cost of spatial jump, while ensuring that the preview benefit is not reduced or is increased. Preferably, the above indicators can be combined into a comprehensive score according to business weights, and the candidate sequence segment with the better comprehensive score can be selected. The updated local sequence segments are written back to the uploaded sequence to obtain the updated sequence, and the target uploaded sequence is output when the preset termination conditions are met. The preset termination conditions include at least: the number of iterations reaches a threshold (preferably 5-30 times), the improvement of the overall score is less than a threshold (preferably less than 1-5%), or the optimization is completed within the preset time budget (preferably 10-200ms).

[0033] Specifically, in this embodiment, considering that the upload order of the watermark feature sub-images directly affects the splicing connectivity and progressive preview formation speed of the server, if the upload sequence has frequent segment switching, hierarchical jumps or large spatial jumps, it is easy for the feature distribution obtained by the server at any time to be too discrete, which in turn leads to a large number of disconnected areas in the intermediate feature map and increases the subsequent missing completion overhead.

[0034] To address this, this embodiment introduces a sequence local optimization mechanism on the client side. It constrains the local rearrangement range by using calculable switching costs and spatial jumps, and makes features of the same segment arrive at the server more concentrated in time. At the same time, it makes the hierarchical progression order more compatible with the server's verification and splicing process. Thus, without significantly disturbing the global priority, it can reduce the scheduling overhead during transmission, increase the probability of continuous coverage of intermediate feature maps on the server, thereby improving the stability of progressive preview and reducing the number of iterations for reconstructing residual images.

[0035] In another embodiment, when multiple source images all require the generation of watermark feature sub-images and uploading to the server, the client, after generating initial upload sequences corresponding to the multiple source images respectively, performs cross-source image sequence secondary adjustment processing on the multiple initial upload sequences to generate a joint target upload sequence, specifically including: Obtain multiple initial upload sequences corresponding to multiple source images, and label each initial upload sequence with a source image identifier, wherein the source image identifier is preferably a file identifier, task identifier, or session identifier; A cross-source scheduling state set is determined, wherein the cross-source scheduling state set refers to the set of runtime state parameters required for cross-source scheduling, including at least transmission state parameters, server confirmation state parameters, and source image completion parameters; preferably, the server confirmation state parameters can be obtained through server receipts or confirmation messages, and at least represent one or more of the following: the number of received items, the missing item list, and the most recent confirmation time; the source image completion parameters are preferably obtained through the "ratio of the number of uploaded items to the total number of items", and optionally further obtained through the "ratio of items that have met the preview threshold". Determine the minimum previewable coverage constraint for each of the multiple source images; wherein, the minimum previewable coverage constraint refers to the minimum arrival condition that the server must meet to generate the initial preview of the source image; preferably, the minimum previewable coverage constraint can be set as: the basic level data reaches at least a number of key segments (e.g., 4, 8 or 16), or the basic level coverage ratio reaches a preset threshold (e.g., 5-20%); wherein, the key segments are preferably central area segments, business-focused area segments, or a priority set preset according to segment identifier; The cross-source rotation quota set is determined based on the minimum previewable coverage constraint; wherein, the cross-source rotation quota set refers to the set of limiting parameters for each source image to occupy uploaded resources within a time window, including at least the maximum continuous occupation threshold and the maximum waiting latency threshold; preferably, the maximum continuous occupation threshold can be set as the upper limit of the number of consecutive upload items (e.g., 10-50 items) or the upper limit of the continuous occupation duration (e.g., 200-2000ms); the maximum waiting latency threshold can be set as 500-5000ms; Set up a cross-source concurrent slot set and map the upload entries of multiple source images to the cross-source concurrent slot set. The concurrent slot set refers to the set of logical channels used to limit the number of cross-source concurrent uploads. The number of slots is preferably 1-8. Preferably, the number of slots can be adaptive with the transmission status: increase the number of slots to improve throughput when the link is stable, and reduce the number of slots to reduce switching costs when the link fluctuates. A cross-source scheduling weight set is generated based on the cross-source scheduling state set and the cross-source round-robin quota set. The next item to be uploaded is selected from multiple initial upload sequences according to the cross-source scheduling weight set to construct a joint target upload sequence. Specifically, the generation logic of the cross-source scheduling weight can be described as follows: increase the weight of source images that have not yet met the minimum previewable coverage constraint; increase the weight of source images whose waiting latency is close to the threshold; increase the weight of the corresponding missing items of source images that have many missing items reported by the server and need to be filled. In response to transmission status parameters meeting preset jitter or packet loss conditions, a cross-source freeze strategy is triggered. The preset jitter condition can be set to round-trip latency fluctuation exceeding a preset threshold or jitter exceeding a preset threshold, for example, fluctuation reaching the level of hundreds of milliseconds or jitter reaching the level of tens of milliseconds. The preset packet loss condition can be set to packet loss level exceeding a preset proportion threshold, for example, exceeding 3-5%. The cross-source freeze strategy refers to the strategy of reducing the cross-source switching frequency and increasing the proportion of continuous uploads from the same source during the freeze period. The freeze period is preferably 200-2000ms. After the freeze period ends, a cross-source unfreezing strategy is triggered to restore cross-source rotation. Output the joint target upload sequence and upload watermark feature sub-images corresponding to multiple source images accordingly.

[0036] Specifically, in this embodiment, the core challenge of multi-source concurrent uploads lies in the resource competition between different source images and the desire of each source image to meet the server's initial preview conditions as quickly as possible. To address this, this embodiment prioritizes each source image to reach the preview threshold through a minimum previewable coverage constraint, and avoids a single source image from occupying upload resources for a long time through a cross-source rotating quota set. At the same time, it utilizes a cross-source freeze and unfreeze strategy to suppress the additional confirmation and retransmission overhead caused by frequent cross-source switching during network jitter, thereby improving the first frame preview formation speed and overall upload stability of multi-source images.

[0037] In one embodiment, such as Figure 3 As shown, an image transmission method is provided, which is applied to... Figure 1 Taking server-side error 104 as an example, the following steps are included: Step 301: Receive multiple target feature sub-images uploaded by the client, and determine the feature level corresponding to the multiple target feature sub-images and the segmentation identifier of the corresponding segmented image based on the description information associated with the target feature sub-images. The description information refers to the structured metadata carried with or reported separately from the target feature sub-image, including at least one or more of the following: source image identifier, piece identifier, feature level identifier, spatial location description, encoding method identifier, and verification summary; preferably, the description information can be generated by the client based on the piece association information and scheduling entries. Step 302: Collect multiple target feature sub-images based on feature levels, and perform splicing processing on the collected multiple target feature sub-images according to the spatial relative position relationship indicated by the segmentation identifier to generate intermediate feature images corresponding to multiple feature levels. Here, aggregation refers to grouping according to feature level (and optional source image identifiers); stitching refers to placing each feature sub-image into a unified feature canvas according to its spatial location description; for overlapping regions, average fusion, weighted fusion, or confidence-based retention strategies are preferred. Step 303: Based on preset integrity conditions, perform image integrity verification on multiple intermediate feature maps; The preset integrity conditions include at least one or more of the following: coverage ratio threshold, connectivity threshold, and key area integrity threshold; preferably, the coverage ratio threshold can be set to 60-95%; the connectivity threshold can be set to "the maximum connectivity coverage accounts for more than 70% of the covered area"; the key area is preferably the central area, the business focus area, or a preset set of key segments; specifically, the coverage ratio can be obtained by statistically analyzing the proportion of filled positions to the total positions, and the connectivity can be obtained by connected domain analysis; Step 304: In response to one or more intermediate feature maps failing the image integrity check, the intermediate feature maps are compared across multiple feature levels to generate a stitching comparison result. Based on the stitching comparison result, missing inference and incomplete image reconstruction processing are performed on the intermediate feature maps that failed the check to fill in the missing areas and update the intermediate feature maps until all intermediate feature maps pass the image integrity check. Among them, the splicing and comparison results preferably include missing mask, missing block list and missing location range description; Step 305: Perform cross-feature level fusion processing on the multiple intermediate feature maps that have passed the verification to generate a fused feature map, and generate the target image based on the fused feature map.

[0038] In the above image transmission method, the server collects and stitches the uploaded target feature sub-images based on feature level and fragment identifier. When the integrity check fails, the server uses the correspondence between different feature levels to perform missing inference and fragment reconstruction, so that the target image can be gradually restored even if the fragments are incomplete or the arrival order is uncertain. This makes the image generation process no longer dependent on the client to upload the complete image at once, but supports progressive reconstruction based on the received feature data, which significantly improves the engineering feasibility of image transmission and generation in unstable network environments.

[0039] In one embodiment, multiple intermediate feature maps are compared across multiple feature levels to generate a spliced ​​comparison result. Based on the spliced ​​comparison result, missing inference and incomplete image reconstruction processing are performed on the intermediate feature maps that fail the verification to fill in the missing regions and update the intermediate feature maps, including: Obtain watermark feature sub-images corresponding to multiple feature levels, and perform splicing processing on each feature level to generate intermediate feature images. Here, the integrity verification result refers to the validity judgment result output by the existing splicing verification process on the server; the target intermediate feature image refers to the intermediate feature image that does not meet the passing condition. Determine a reference hierarchy set, wherein the reference hierarchy set refers to the set of feature hierarchy levels used to provide reference information for missing completion, preferably including the levels that have passed the verification and the levels with higher integrity scores; Cross-level alignment is performed on the target intermediate feature map and the reference intermediate feature map in the reference level set; where cross-level alignment refers to the process of mapping features from different levels to a unified coordinate domain; preferably, the alignment can be obtained by: first, transforming the reference level features to the target level resolution proportionally according to the resolution reduction relationship of each level; then aligning the two to the same feature canvas coordinates according to the piecewise spatial position description; when there is geometric offset, affine registration or keypoint-based registration can be optionally introduced to reduce the deviation; Based on the cross-level alignment results, difference detection processing is performed to generate splicing and comparison results. The difference detection processing can be obtained as follows: first, "covered area descriptions" of the target level and the reference level are generated separately, and then the two are compared in a unified coordinate domain. The parts "covered by the reference level but not by the target level" are marked as missing areas, thereby obtaining a missing mask and a set of missing blocks. Optionally, the positions where the difference in the aligned feature values ​​exceeds a preset amplitude threshold can also be marked as abnormal missing candidates. The amplitude threshold is preferably configured according to historical statistics or experience, such as taking the high quantile threshold of the feature difference distribution. Missing inference processing is performed based on the splicing comparison results; wherein, missing inference processing refers to the process of generating prior features for the missing region using the reference level; preferably including: mapping the features of the reference level at the corresponding position of the missing region to the target level scale; and interpolating and smoothing the prior features within the missing region based on the continuity of the surrounding spliced ​​regions; at the same time, introducing boundary consistency constraints at the missing boundary to make the missing filling region consistent with the existing region in boundary transition. Residual image reconstruction is performed based on prior features and boundary features to generate missing parts and update the target intermediate feature map; wherein, residual image reconstruction refers to the process of merging the missing parts with the spliced ​​regions, preferably filling in the missing regions with inference features, and using gradient blending or multi-scale fusion at the boundaries to suppress splicing seams. When the preset iteration termination conditions are met, the updated intermediate feature map is output; wherein, the preset iteration termination conditions include at least: the number of iterations reaches a threshold (preferably 1-10 times), the integrity verification result meets the pass condition, or the reduction in the area of ​​the missing region is less than a threshold (preferably less than 1-5%).

[0040] Specifically, in this embodiment, due to factors such as network packet loss or discrete uploaded sequences, the server may experience local missing features at a certain feature level, resulting in disconnected intermediate feature maps and affecting the quality of subsequent target image generation. To address this, this embodiment uses "the existence of intermediate feature maps that have failed integrity verification" as the triggering condition, obtains a missing mask based on cross-level alignment and difference detection, and utilizes the stable coverage of the reference level to provide prior features for the missing region. At the same time, boundary consistency constraints are introduced to reduce the risk of discontinuity between the completed region and the stitched region, thereby reducing the reconstruction uncertainty caused by single-level missing features and improving the integrity of the intermediate feature maps and the stability of subsequent generation results.

[0041] In one embodiment, performing cross-feature-level fusion processing on multiple verified intermediate feature maps to generate a fused feature map includes: Multiple intermediate feature maps are obtained and the feature level identifiers corresponding to each intermediate feature map are determined. The feature level identifiers refer to the level labels used to characterize the feature granularity and / or semantic level. Scale unification and coordinate alignment are performed on multiple intermediate feature maps to generate an aligned multi-level feature set. Scale unification means transforming the features of each level to the same target resolution by upsampling or downsampling at the same ratio, and coordinate alignment means placing the features of each level into the same feature canvas coordinate domain according to the segmented spatial location description. Determine the fusion weight set; where the fusion weight set refers to the set of weight parameters used to allocate the contribution of different feature levels; specifically, the fusion weight can be obtained in the following way: calculate the integrity score, coverage, connectivity, and watermark semantic relevance of each level respectively, and weight the above indicators according to preset coefficients to obtain the level score, and then normalize the scores of each level to obtain the fusion weight; preferably, the preset coefficient can be set as follows: "integrity weight is high, coverage is second, connectivity is third, and semantic relevance is low", for example, it can be configured in the proportion of 0.4, 0.3, 0.2, 0.1, and can be adjusted according to business; A fusion operation is performed on the aligned multi-level feature set to generate a fused feature map; wherein, the fusion operation refers to the operation process of combining multi-level features to obtain a unified feature representation, preferably including one or more of the following: weighted superposition of features at each level according to the fusion weight, adaptive selection of more reliable level features at different spatial locations by a gating strategy and combination, and reduction of the contribution of levels with more missing features and supplementation by levels with more complete coverage. Consistency constraint processing is performed on the fused feature map; wherein, consistency constraint processing refers to the processing used to suppress fusion artifacts and enhance cross-level structural consistency, including at least one or more of edge consistency constraints, structural similarity constraints and noise suppression processing; preferably, the edge / gradient changes of the fusion result can be smoothed by applying constraints, or the fusion result can be guided to maintain structural stability by evaluating structural similarity; Output a fused feature map and generate a target image based on the fused feature map; wherein, the target image is preferably generated by decoding the fused feature map through a preset decoding network, a generation network or a reconstruction network.

[0042] Specifically, in this embodiment, intermediate feature maps at different feature levels are usually complementary in terms of global structure preservation and local detail representation. If the target image is generated by relying on only a single level, it is easy to have a stable structure but insufficient detail or rich detail but structural drift. To this end, this embodiment achieves multi-level fusion conditions through scale unification and coordinate alignment, and adaptively determines the fusion weight based on the integrity verification score and coverage index, thereby improving the detail representation ability while ensuring structural consistency. At the same time, the consistency constraint suppresses the edge discontinuity and noise amplification problems caused by fusion, so as to improve the visual consistency and robustness of the final target image.

[0043] In one specific embodiment, during the transmission of multiple target feature sub-images between the client and the server, the method further includes: the server sending a receipt confirmation message and the client confirming the upload sequence progress in real time to support breakpoint resumption and missing data replacement, specifically including the following steps: After receiving any target feature sub-image, the server generates a reception confirmation message and sends it to the client. This reception confirmation message indicates that the server has successfully received and stored the target feature sub-image on disk or cached it. Preferably, the reception confirmation message includes at least one or more of the following: source image identifier, slice identifier, feature level identifier, reception status identifier, reception timestamp, data verification result, and sequence progress identifier. The data verification result is preferably obtained by performing a verification digest comparison or error correction verification on the received data, and is used to determine data integrity and consistency. The client receives the confirmation information and updates the upload sequence progress based on the confirmation information; wherein, the upload sequence progress refers to the client's status record for the "set of target feature sub-graphs that have been confirmed to have been uploaded" and the "set of target feature sub-graphs to be uploaded"; preferably, the client maintains an upload status flag for each target feature sub-graph, and the upload status flag includes at least one or more of the following: not uploaded, sent but not confirmed, confirmed, and needs to be retransmitted. The server sends batch confirmation information when the preset confirmation trigger conditions are met; wherein, the preset confirmation trigger conditions include at least one or more of the following: the cumulative number of received entries reaches a threshold, the cumulative amount of received data reaches a threshold, the preset time interval is reached or link jitter enhancement is detected; the batch confirmation information preferably represents the received set in the form of a confirmation bitmap, a confirmation list or a confirmation interval to reduce confirmation overhead; When the server detects a sequence gap, it generates a missing item notification and sends it to the client. The sequence gap refers to an item that has not been received or has failed verification in the set of items that the uploaded sequence should have reached. The missing item notification includes at least a description of the set of missing items or the range of missing locations. Preferably, the set of missing items can be merged according to the fragment identifier and the feature level identifier, and can optionally indicate the reason for the missing items (e.g., timeout, verification failure, or version mismatch). The client performs breakpoint resume processing based on the missing message; breakpoint resume processing refers to locating unconfirmed items in the upload sequence progress and retransmitting or supplementing them. Specifically, the client can perform the following actions: add the set of missing entries to the high-priority retransmission queue, and prioritize retransmitting the missing entries without changing the global upload strategy; preferably, when the transmission status is poor, prioritize retransmitting the corresponding entries of the basic level or key fragments so that the server can restore the connectivity of the intermediate feature map as soon as possible. When the client detects a transmission interruption or session reconstruction, the client sends a progress query request to the server; the server returns the current confirmed set; the client then resumes the upload sequence progress and continues uploading unconfirmed items; wherein, session reconstruction preferably includes network reconnection, application restart, or task resumption scenarios; the confirmed set preferably adopts a bitmap / list / range representation format consistent with batch confirmation.

[0044] Specifically, in this embodiment, the server confirms the received target feature sub-graphs and sends a sequence gap feedback to the client, enabling the client to monitor the upload progress in real time and retransmit items that have not arrived or failed verification. This allows for breakpoint resumption in the event of network interruption, jitter, or client restart, avoiding the additional bandwidth consumption and latency increase caused by retransmitting the entire sequence, and improving the reliability and controllability of the multi-target feature sub-graph transmission process.

[0045] It should be understood that, although Figure 2 and Figure 3 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 2 and Figure 3 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0046] In one embodiment, such as Figure 4 As shown, an image transmission device is provided, which includes a client and a server. The device specifically includes: Clients, including: The slicing module acquires the source image and performs slicing processing on the source image to obtain multiple sliced ​​images; The watermark generation module generates watermark information data for multiple segmented images based on image recognition information and spatiotemporal parameters of the client, and generates watermark carrier images for multiple segmented images based on the watermark information data and preset watermark configuration parameters. The layering module extracts features from the sliced ​​image and the watermark carrier image based on multiple preset feature levels, generating multiple slice feature sub-images and multiple watermark carrier feature sub-images respectively. The fusion module determines the watermark mark position of multiple watermark carrier feature sub-images based on the relative positions of multiple segmented feature sub-images in the source image, so as to add the multiple watermark carrier feature sub-images to the corresponding positions of the multiple segmented feature sub-images and fuse them to generate multiple target feature sub-images. The upload module determines the upload sequence of multiple target feature sub-images based on the transmission status information from the client to the server, the watermark information of the target feature sub-image, and the feature level corresponding to the target feature sub-image, and uploads the multiple target feature sub-images to the server based on the upload sequence; The server side includes: The receiving module is used to receive multiple target feature sub-images uploaded by the client, and determine the feature level corresponding to the multiple target feature sub-images and the segment identifier of the corresponding segmented image based on the description information associated with the target feature sub-images. The stitching module is used to aggregate multiple target feature sub-images based on feature levels, and to stitch the aggregated target feature sub-images according to the spatial relative positional relationship indicated by the piece identifier, generating intermediate feature maps corresponding to multiple feature levels. The verification module is used to perform image integrity verification on multiple intermediate feature maps based on preset integrity conditions. The completion module is used to compare multiple intermediate feature maps across multiple feature levels in response to one or more intermediate feature maps failing the image integrity check, generate a stitching comparison result, and perform missing inference and incomplete image reconstruction processing on the intermediate feature maps that failed the check based on the stitching comparison result, so as to complete the missing regions and update the intermediate feature maps until multiple intermediate feature maps all pass the image integrity check. The generation module is used to perform cross-feature level fusion processing on multiple intermediate feature maps that have passed the verification, generate a fused feature map, and generate the target image based on the fused feature map.

[0047] Specific limitations regarding the image transmission device can be found in the limitations of the image transmission method described above, and will not be repeated here. Each module in the aforementioned image transmission device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0048] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements an image transmission method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0049] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0050] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: The source image is acquired and sliced ​​to obtain multiple sliced ​​images; Based on the image recognition information of multiple image segments and the spatiotemporal parameters of the client, watermark information data of multiple image segments is generated, and watermark carrier images of multiple image segments are generated based on the watermark information data and preset watermark configuration parameters. Based on multiple preset feature levels, feature layer extraction is performed on the sliced ​​image and the watermark carrier image respectively to generate multiple sliced ​​feature sub-images and multiple watermark carrier feature sub-images. The watermark mark positions of multiple watermark carrier feature sub-images are determined based on the relative positions of multiple segmented feature sub-images in the source image, so that the multiple watermark carrier feature sub-images are added to the corresponding positions of multiple segmented feature sub-images and fused to generate multiple target feature sub-images. Based on the transmission status information from the client to the server, the watermark information of the target feature sub-image, and the feature level corresponding to the target feature sub-image, the upload sequence of multiple target feature sub-images is determined, and the multiple target feature sub-images are uploaded to the server based on the upload sequence.

[0051] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: The source image is acquired and sliced ​​to obtain multiple sliced ​​images; Based on the image recognition information of multiple image segments and the spatiotemporal parameters of the client, watermark information data of multiple image segments is generated, and watermark carrier images of multiple image segments are generated based on the watermark information data and preset watermark configuration parameters. Based on multiple preset feature levels, feature layer extraction is performed on the sliced ​​image and the watermark carrier image respectively to generate multiple sliced ​​feature sub-images and multiple watermark carrier feature sub-images. The watermark mark positions of multiple watermark carrier feature sub-images are determined based on the relative positions of multiple segmented feature sub-images in the source image, so that the multiple watermark carrier feature sub-images are added to the corresponding positions of multiple segmented feature sub-images and fused to generate multiple target feature sub-images. Based on the transmission status information from the client to the server, the watermark information of the target feature sub-image, and the feature level corresponding to the target feature sub-image, the upload sequence of multiple target feature sub-images is determined, and the multiple target feature sub-images are uploaded to the server based on the upload sequence.

[0052] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0053] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0054] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. An image transmission method, characterized in that, Applied to the client side, including: Acquire a source image and perform segmentation processing on the source image to obtain multiple segmented images; Based on the image recognition information of multiple segmented images and the spatiotemporal parameters of the client, watermark information data of multiple segmented images is generated, and watermark carrier images of multiple segmented images are generated based on the watermark information data and preset watermark configuration parameters. Based on multiple preset feature levels, feature layer extraction is performed on the segmented image and the watermark carrier image respectively to generate multiple corresponding segmented feature sub-images and multiple watermark carrier feature sub-images. Based on the relative positions of the multiple segmented feature sub-images in the source image, the watermark mark positions of the multiple watermark carrier feature sub-images are determined, so that the multiple watermark carrier feature sub-images are added to the corresponding positions of the multiple segmented feature sub-images and fused to generate multiple target feature sub-images. Based on the transmission status information from the client to the server, the watermark information of the target feature sub-image, and the feature level corresponding to the target feature sub-image, an upload sequence of multiple target feature sub-images is determined, and multiple target feature sub-images are uploaded to the server based on the upload sequence.

2. The image transmission method according to claim 1, characterized in that, The process of acquiring a source image and performing segmentation on the source image to obtain multiple segmented images includes: The image size information of the source image is obtained, and preprocessing is performed on the source image based on the image size information to obtain a preprocessed image. The preprocessing includes at least one or more of the following: scaling, cropping, padding, and normalization. Based on the preset segmentation size and the overlap ratio of the segmented images, multiple segmentation regions are determined in the preprocessed image; Based on the multiple segmented regions, multiple corresponding segmented images are cropped from the preprocessed image, and the segmentation identifiers of the multiple segmented images are determined according to the cropping order. Based on the position parameters of the segmented image in the preprocessed image and the segment identifier, the segment association information of the segmented image is generated.

3. The image transmission method according to claim 1, characterized in that, The step of generating watermark information data for multiple image segments based on image recognition information of multiple image segments and spatiotemporal parameters of the client includes: Based on a preset image recognition model, image recognition processing is performed on multiple sliced ​​images to obtain image recognition information of multiple sliced ​​images. The image recognition information includes at least one or more of the following: object category, scene label, text recognition result, and object position parameters. Obtain the spatiotemporal parameters of the client, at least the time parameter and the geographical location parameter; Determine the categories of hot business objects and the boundary information of hot areas based on preset business association parameters; Based on the geographic location parameters and the object location parameters, the location information of the object in the segmented image is fused and determined to obtain object location association information; The object location association information is compared with the hotspot region boundary information to determine whether the object in the segmented image is located in the preset hotspot region, and hotspot region discrimination information is generated. The object category is compared with the business hotspot object category to determine whether the object in the segmented image is a business hotspot object, and hotspot object discrimination information is generated; One or more of the image recognition information, spatiotemporal parameters, hotspot region discrimination information, and hotspot object discrimination information of the segmented image are combined to determine the watermark payload of the segmented image, and the watermark payload is encoded to generate the watermark information data of the segmented image.

4. The image transmission method according to claim 1, characterized in that, The step of extracting features from the segmented image and the watermark carrier image based on multiple preset feature levels to generate multiple corresponding segmented feature sub-images and multiple watermark carrier feature sub-images includes: A first feature extraction network is constructed based on multiple preset feature levels, and the segmented image is input into the first feature extraction network; Based on the feature extraction path in the first feature extraction network, multiple convolution operations are performed on the segmented image, and after at least one convolution operation, downsampling operation is performed on the feature map obtained by the convolution operation to generate multiple segmented feature sub-maps corresponding to the segmented image, and the multiple segmented feature sub-maps correspond one-to-one with the multiple feature levels; A second feature extraction network is constructed based on multiple feature levels consistent with the first feature extraction network; The watermark carrier image is subjected to frequency domain transformation to obtain a frequency domain coefficient map, and the frequency domain coefficient map is input into the second feature extraction network; Based on the feature extraction path in the second feature extraction network, a convolution operation is performed on the frequency domain coefficient map, and pooling and / or downsampling operations are performed on the feature map obtained by the convolution operation to generate multiple watermark intermediate images corresponding to the watermark carrier image. The multiple watermark intermediate images correspond to multiple feature levels. Scale alignment processing is performed on multiple watermark intermediate images to generate multiple watermark carrier feature sub-images corresponding to the watermark carrier image, so that the multiple segment feature sub-images corresponding to the segmented images correspond one-to-one with the multiple watermark carrier feature sub-images corresponding to the watermark carrier image at multiple feature levels.

5. The image transmission method according to claim 1, characterized in that, The step of determining the watermark mark position of multiple watermark carrier feature sub-images based on their relative positions in the source image includes: Obtain the segmentation position information of multiple segmented images in the source image, and determine the relative positional relationship of multiple segmented feature sub-images in the source image based on the segmentation position information; Based on the relative positional relationship, the spatial mapping relationship between the segmented feature sub-image and the source image is hierarchically aligned at multiple feature levels to generate hierarchical positional mapping results corresponding to multiple feature levels; Based on the hierarchical position mapping result, the candidate watermark mark positions of multiple watermark carrier feature sub-images on the source image are determined, so that the relative position relationship between the candidate watermark mark positions of multiple watermark carrier feature sub-images and multiple segmented feature sub-images remains consistent. Boundary constraints and / or conflict resolution processes are performed on multiple candidate watermark locations to generate multiple watermark location sub-maps of the watermark carrier.

6. The image transmission method according to claim 1, characterized in that, The step of determining an upload sequence for multiple target feature sub-images based on the transmission status information from the client to the server, the watermark information of the target feature sub-image, and the feature level corresponding to the target feature sub-image, and uploading the multiple target feature sub-images to the server based on the upload sequence, includes: Obtain and parse the transmission status information from the client to the server to obtain transmission association parameters; A transmission status index is generated based on the transmission association parameters to at least characterize the transmission status level from the client to the server; The watermark information of the target feature sub-image is obtained and parsed to obtain the watermark payload elements; Semantic recognition is performed on the watermark payload elements to generate semantic recognition results, and watermark semantic indicators are determined based on the correlation between the semantic recognition results and preset key semantics. Based on the feature hierarchy of the target feature sub-image, a mapping is performed to determine the feature hierarchy index of the target feature sub-image, wherein the feature hierarchy index is positively correlated with the abstract scale level of the target feature sub-image; A first weight is obtained by performing a positive mapping based on the transmission status index, so as to at least characterize the contribution of the watermark semantic index to the upload priority. A second weight is obtained by performing a reverse mapping based on the transmission status index, so as to at least characterize the contribution of the feature level index to the upload priority. Based on the first weight, the second weight, the watermark semantic index, and the feature level index, the upload priority index of the target feature sub-image is determined; The target feature sub-graphs are sorted according to the multiple upload priority indicators to generate an initial upload sequence; Based on preset sequence local optimization rules, the initial upload sequence is subjected to sequence local optimization processing to generate a target upload sequence, and multiple target feature sub-images are uploaded to the server based on the target upload sequence.

7. An image transmission method, characterized in that, Applied to the server side, the method includes: Receive multiple target feature sub-images uploaded by the client, and determine the feature level corresponding to the multiple target feature sub-images and the segment identifier of the corresponding segmented image based on the description information associated with the target feature sub-images; Based on the feature hierarchy, multiple target feature sub-images are aggregated, and according to the spatial relative position relationship indicated by the segmentation identifier, the aggregated multiple target feature sub-images are spliced ​​to generate intermediate feature images corresponding to multiple feature levels; Based on preset integrity conditions, image integrity verification is performed on multiple intermediate feature maps; In response to one or more of the intermediate feature maps failing the image integrity check, the intermediate feature maps are compared across multiple feature levels to generate a stitching comparison result. Based on the stitching comparison result, missing inference and fragment reconstruction processing are performed on the intermediate feature maps that failed the check to fill in the missing regions and update the intermediate feature maps, until all of the intermediate feature maps pass the image integrity check. Perform cross-feature-level fusion processing on multiple intermediate feature maps that have passed the verification to generate a fused feature map, and generate a target image based on the fused feature map.

8. An image transmission device, characterized in that, The apparatus includes a client and a server for implementing the image transmission method as described in any one of claims 1-6, wherein the server includes: The receiving module is used to receive multiple target feature sub-images uploaded by the client, and determine the feature level corresponding to the multiple target feature sub-images and the segment identifier of the corresponding segmented image based on the description information associated with the target feature sub-images. The stitching module is used to aggregate multiple target feature sub-images based on the feature levels, and to perform stitching processing on the aggregated multiple target feature sub-images according to the spatial relative position relationship indicated by the segmentation identifier, thereby generating intermediate feature maps corresponding to the multiple feature levels. The verification module is used to perform image integrity verification on multiple intermediate feature maps based on preset integrity conditions; The completion module is configured to, in response to one or more of the intermediate feature maps failing the image integrity check, compare the multiple intermediate feature maps across multiple feature levels, generate a stitching comparison result, and perform missing inference and fragment reconstruction processing on the intermediate feature maps that failed the check based on the stitching comparison result, so as to complete the missing regions and update the intermediate feature maps, until all the intermediate feature maps pass the image integrity check. The generation module is used to perform cross-feature level fusion processing on multiple intermediate feature maps that have passed the verification, generate a fused feature map, and generate a target image based on the fused feature map.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images