Image labeling method, device and computer readable storage medium
By performing image denoising and pixel gradient calculation, combined with secondary annotation technology, the problems of low efficiency and poor accuracy in image annotation are solved, achieving efficient and accurate image annotation and improving the training effect of neural network models.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHIDAO NETWORK TECH (BEIJING) CO LTD
- Filing Date
- 2023-01-09
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies suffer from low efficiency and poor accuracy in image annotation, resulting in high training costs and low recognition accuracy for neural network models.
The image to be labeled is denoised, the pixel gradient of the denoised image is calculated, and secondary labeling is performed based on the pixel gradient to obtain fine boundaries.
It improves the efficiency and accuracy of image annotation, reduces annotation costs, and enhances the recognition accuracy of neural network models.
Smart Images
Figure CN116258895B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of image processing, and more particularly to image annotation methods, devices, and computer-readable storage media. Background Technology
[0002] Semantic segmentation is a pixel-level classification task for images, typically implemented using pre-trained neural network models. Before practical application, semantic segmentation-based neural network models require training with a vast number of image samples. These massive training sets need to be labeled. Currently, image labeling primarily relies on manual annotation. However, manual image labeling is not only inefficient but also inaccurate. The resulting samples are costly for training neural network models, and the final model often exhibits poor recognition accuracy. Summary of the Invention
[0003] To address or partially address the problems existing in related technologies, this application provides an image annotation method, device, and computer-readable storage medium, which can improve the efficiency and accuracy of image sample annotation.
[0004] The first aspect of this application provides an image annotation method, including:
[0005] The image to be labeled is denoised to obtain a denoised image;
[0006] Calculate the pixel gradient of the denoised image;
[0007] Based on the pixel gradient of the denoised image, secondary annotation is performed on each target to be labeled in the denoised image to obtain the fine boundary of each target to be labeled.
[0008] A second aspect of this application provides an image annotation apparatus, comprising:
[0009] The denoising module is used to denoise the image to be labeled, so as to obtain a denoised image.
[0010] A calculation module is used to calculate the pixel gradient of the denoised image;
[0011] The secondary annotation module is used to perform secondary annotation on each target to be annotated in the denoised image based on the pixel gradient of the denoised image, so as to obtain the fine boundary of each target to be annotated.
[0012] A third aspect of this application provides an electronic device, comprising:
[0013] Processor; and
[0014] A memory that stores executable code, which, when executed by the processor, causes the processor to perform the method described above.
[0015] A fourth aspect of this application provides a computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method described above.
[0016] The technical solution provided in this application can include the following beneficial effects: Unlike related technologies that use manual annotation when annotating image samples, resulting in high annotation costs and low accuracy, the technical solution of this application first performs denoising processing on the image to be annotated. The denoised image is used as the basis for calculation, avoiding the adverse effects of noise in the original image to be annotated on the accuracy of subsequent calculations. On the other hand, after calculating the pixel gradient of the denoised image, secondary annotation is performed on each target to be annotated in the denoised image based on the pixel gradient of the denoised image. Since the secondary annotation is based on at least one manual annotation, automatic or semi-automatic annotation, the boundary is further adjusted towards the pixel direction of the gradient increase (i.e., the position where the color change is most drastic), making the boundary more accurate. Therefore, not only is the annotation efficiency high, but the annotation accuracy is also high.
[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description
[0018] The above and other objects, features and advantages of this application will become more apparent from the following description of exemplary embodiments of this application in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components.
[0019] Figure 1 This is a flowchart illustrating the image annotation method provided in an embodiment of this application;
[0020] Figure 2 The image to be labeled includes target objects such as riders, vehicles, trees, flowers and road signs, as provided in the embodiments of this application.
[0021] Figure 3a The embodiments provided in this application are for Figure 2 A schematic diagram illustrating the results of manual annotation of the example image to be annotated;
[0022] Figure 3b The embodiments provided in this application are based on Figure 3a The example of manually annotated results is a schematic diagram of the first boundary of each target to be annotated in a denoised image;
[0023] Figure 4 This is a schematic diagram of the image annotation device provided in the embodiments of this application;
[0024] Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation
[0025] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.
[0026] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.
[0027] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.
[0028] Semantic segmentation is a pixel-level classification task for images, typically implemented using neural network models. Before practical application, semantic segmentation-based neural network models require training with a vast number of image samples. These massive training sets need to be labeled. Currently, image labeling is primarily done manually. However, manual labeling is not only inefficient but also inaccurate. The resulting samples are costly for training neural network models, and the final model often exhibits poor recognition accuracy.
[0029] To address the aforementioned issues, this application provides an image annotation method that can improve the efficiency and accuracy of image sample annotation.
[0030] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.
[0031] See Figure 1 This is a flowchart illustrating the image annotation method in an embodiment of this application, mainly including steps S101 to S103, as described below:
[0032] Step S101: Denoise the image to be labeled to obtain a denoised image.
[0033] An image to be labeled is an image containing some target objects (e.g., people, vehicles, flowers, trees, birds, animals, insects, fish, etc.). Since these images are to be used as a training set to train a neural network, it is necessary to label the target objects they contain. Figure 2 As shown, the rider, vehicles, and traffic signs in the image are the target objects in the image to be labeled. Since some noise is inevitably introduced when capturing the image to be labeled, the accuracy of subsequent calculations (e.g., calculating pixel gradients) will be greatly reduced if this noise is not removed. Therefore, in this embodiment, the image to be labeled can be denoised to obtain a denoised image. For example, Gaussian blur, mean filtering, and other techniques can be used to denoise and smooth the image to obtain a denoised image. It should be noted that both Gaussian blur and mean filtering use convolution kernels to convolve the image to be labeled. The difference is that the former's convolution kernel generally conforms to a Gaussian distribution, while the latter's convolution kernel has all elements with equal values.
[0034] As one embodiment of this application, denoising the image to be labeled to obtain a denoised image may involve: determining a target mean square error image with the same size as the image to be labeled; determining a target Gaussian blur kernel corresponding to each pixel in the image to be labeled based on a preset blur radius and the target mean square error image; and for each pixel in the image to be labeled, determining the target Gaussian blur value corresponding to the pixel based on the target Gaussian blur kernel and target pixels around the pixel within a preset blur radius, so as to obtain a Gaussian blurred image corresponding to the image to be labeled. It should be noted that the target mean squared error image mentioned above is called a mean squared error image because the pixel values of the pixels in it are used to determine the mean squared error of the target Gaussian blur kernel for the corresponding pixel in the image to be labeled. This can be determined as follows: First, based on the initial mean squared error image, determine the initial Gaussian blur kernel corresponding to each pixel in the image to be labeled. Then, for each pixel in the image to be labeled, determine the initial Gaussian blur value corresponding to the pixel based on the pixel values of the surrounding pixels and the initial Gaussian blur kernel, thus obtaining the initial Gaussian blur image corresponding to the image to be labeled. Next, input the initial Gaussian blur image into the trained image processing model to calculate the target loss function. Finally, update the initial mean squared error image based on the calculation result of the target loss function to obtain the target mean squared error image.
[0035] In another embodiment of this application, denoising the image to be labeled to obtain a denoised image may involve: performing mean filtering on the image to be labeled and the corresponding reference image to construct a first noise characteristic map; performing downsampling and mean filtering on the target color channel of the image to be labeled and the corresponding reference image to construct a second noise characteristic map; determining the noise intensity corresponding to the image to be labeled by combining the first noise characteristic map and the brightness features corresponding to the image to be labeled; and determining a target noise characteristic map based on the noise intensity combined with the first and second noise characteristic maps, so as to remove noise from the image to be labeled according to the target noise characteristic map. In the above embodiment, the reference image corresponding to the image to be labeled may be several frames of images that are consecutive to and preceding the image to be labeled, and the format may be a grayscale image. When constructing the first noise characteristic map, the following steps can be taken: First, calculate the corresponding pixel mean filtering results and pixel block mean filtering results according to a preset window size for the image to be labeled, the first control image, and the second control image; calculate the differences between the pixel mean filtering results of the image to be labeled and the first control image and the second control image, and fuse them based on the differences to determine the first pixel filtering result; calculate the differences between the pixel block mean filtering results of the image to be labeled and the first control image and the second control image, and fuse them based on the differences to determine the first pixel block filtering result; finally, fuse the first pixel filtering result and the first pixel block filtering result to determine the first noise characteristic map.
[0036] Step S102: Calculate the pixel gradient of the denoised image.
[0037] Step S102 can be implemented based on the Sobel operator, that is, calculating the pixel gradient of the denoised image can be based on the Sobel algorithm. In the field of image processing, the Sobel operator is a function with a size of (2^35)^35. k +1)*(2 k +1) convolution kernel, where, k It is a positive integer. For a convolution kernel of (2... k +1)*(2 k The Sobel operator with +1 has the following internal element values:
[0038]
[0039]
[0040] in, Co The fixed coefficient is the smallest integer that makes all elements in the Sobel operator integers, i.e., the least common multiple of all denominators. and They can calculate the image to be labeled. x direction and y The operator for directional gradients. For example, when... k =1, Co When = 2, the specific form of the Sobel operator is as follows:
[0041]
[0042]
[0043] Calculating the pixel gradient of a denoised image based on the Sobel operator essentially involves using a convolution kernel of (2... k +1)*(2 k The Sobel operator (+1) performs a convolution operation on the denoised image. Specifically, for the image to be labeled, the coordinates are (...). i , j ), color channel index is c The pixels, first calculate and as follows:
[0044]
[0045]
[0046] Then, according to the formula
[0047]
[0048] Calculated This refers to the pixel gradient of the denoised image across each color channel.
[0049] Step S103: Perform secondary annotation on each target to be labeled in the denoised image based on the pixel gradient of the denoised image to obtain the fine boundary of each target to be labeled.
[0050] Unlike related technologies that rely on manual annotation for image sample labeling, resulting in high annotation costs and low accuracy, the technical solution of this application involves calculating the pixel gradient of the denoised image using the Sobel operator, and then performing secondary annotation on each target to be labeled in the denoised image based on the pixel gradient of the denoised image. Secondary annotation implies that an initial annotation was performed beforehand, for example, manually. Since the initial annotation is manual, it allows for a relatively quick (and efficient) coarse annotation of the image, while secondary annotation is much more precise and efficient. As an embodiment of this application, step S103 can be implemented as follows: based on the pre-annotated data of the image to be labeled, obtain the first boundary of each target to be labeled in the denoised image; perform secondary annotation on each target to be labeled in the denoised image based on the pixel gradient of the denoised image and the first boundary of each target to be labeled, obtaining the second boundary of each target to be labeled as the fine boundary of each target to be labeled. In the above embodiments, the pre-annotated data of the image to be annotated is obtained by manually (usually by a dedicated annotator) performing semantic segmentation on the image to be annotated, thereby labeling the target objects within it. Specifically, this can be done by filling in color channels or other indexes within the boundaries of the target objects to distinguish the categories of the target objects. Figure 3a As shown, it is for Figure 2 The example image to be labeled shows the results obtained through manual annotation. Through manual annotation, Figure 2 The target objects, such as trees, riders, vehicles, lawns, and traffic signs, are all distinguished by different colors. As for... Figure 3b This is an example of obtaining the first boundary of each target to be labeled in a denoised image, specifically using the Sobel operator. Figure 3a By performing a convolution operation on the manually labeled image in the example, the first boundaries of each unlabeled target in the denoised image can be obtained. It should be noted that, due to the characteristics of the Sobel operator, only the pixel values at the boundary positions are non-zero; therefore, the boundaries of the target objects appear white.
[0051] Further, in the above embodiments, the secondary annotation of each target in the denoised image based on the pixel gradient of the denoised image and the first boundary of each target to be labeled, to obtain the second boundary of each target as the fine boundary of each target, can be achieved by moving the first boundary of each target in the direction of increasing pixel gradient until it reaches the maximum value of the pixel gradient, and using the second boundary of each target obtained at the point of stopping as the fine boundary of each target. As mentioned above, the first boundary of each target is obtained by manually annotating and then performing a convolution operation on the denoised image using the Sobel operator. Since the pre-annotated data is obtained by manual annotation, there is a defect of inaccurate annotation. To address this defect, the technical solution of this application can perform secondary annotation of each target in the denoised image based on the pixel gradient of the denoised image and the first boundary of each target to be labeled, to obtain the second boundary of each target as the fine boundary of each target. Specifically, the first boundary of each target is moved in the direction of increasing pixel gradient until it reaches the maximum value of the pixel gradient, and the second boundary of each target obtained at the point of stopping is used as the fine boundary of each target. In the above embodiments, the location of the maximum pixel gradient indicates the most drastic color change at that position; therefore, the boundary at this location should be the fine boundary of the target to be labeled. It should be noted that, in order to quickly move to the location of the maximum pixel gradient, when moving the first boundary of each target to be labeled in the direction of increasing pixel gradient, a gradient ascent method can be used to move the first boundary of each target to be labeled in the direction of increasing pixel gradient. Let the first boundary of any target to be labeled in the denoised image be obtained based on the pre-labeled data of the image to be labeled. The pixel gradient of the denoised image is Then, the first boundary of each target to be labeled is moved in the direction of increasing pixel gradient until it reaches the maximum pixel gradient, meaning that in Based on continuous movement, new Until the value of Evaluate in the following formula reaches its maximum:
[0052]
[0053] Where H and W are the height and width of the denoised image, respectively, and C is the total number of color channels. Calculated for the foregoing embodiments ,and i , j and c As defined in the foregoing embodiments i , j and cThe details are the same as those in the foregoing embodiments, and will not be repeated here.
[0054] As another embodiment of this application, secondary annotation of each target to be labeled in the denoised image based on the pixel gradient of the denoised image to obtain the fine boundary of each target to be labeled can also be achieved through steps S1031 to S1032, as detailed below:
[0055] Step S1031: Determine the next generation cluster centers based on the pixel gradient of the denoised image and the first generation cluster centers determined by obtaining the first boundary of each target to be labeled.
[0056] In this embodiment of the application, in order to avoid the cluster center falling on the boundary with a large gradient, a preset neighborhood can be determined based on the first-generation cluster center, the pixel gradient between all pixels in the preset neighborhood and the first-generation cluster center can be calculated, and the cluster center can be moved to the pixel position with the smallest pixel gradient. The moved cluster center is the second-generation cluster center.
[0057] Step S1032: Calculate the difference between the next-generation cluster center and each pixel in the preset neighborhood, and segment the pixels in the preset neighborhood according to the difference to form approximate pixel clusters.
[0058] In this embodiment, the difference between the next-generation cluster center and each pixel in the preset neighborhood includes differences in color and spatial properties. During pixel clustering, since each pixel is traversed by multiple next-generation cluster centers, each pixel will have multiple distances to surrounding next-generation cluster centers. Taking the next-generation cluster center corresponding to the minimum distance value as the cluster center of that pixel, the pixel can be assigned to the region where the next-generation cluster center corresponding to the minimum distance value is located, forming an approximate pixel cluster; an approximate pixel cluster means a block-shaped region composed of multiple pixels that are highly similar in color, brightness, etc.
[0059] Step S1033: Cluster the approximate pixel clusters according to the pixel information corresponding to the first boundary delineation region identifier and the pixel information in the approximate pixel clusters to obtain the fine boundary of each target to be labeled.
[0060] In this embodiment, the first boundary delineation region identifier can be a mark made on the target object within the boundary delineation region after obtaining the boundaries of each target to be labeled during manual annotation of the image to be labeled, such as a circle, triangle, etc. Specifically, step S1033 can be implemented as follows: determining the target approximate pixel block from multiple approximate pixel blocks; determining the traversal region based on the cluster center of the target approximate pixel block and a preset distance; obtaining the identifier approximate pixel block corresponding to the identifier in the traversal region; clustering the approximate pixel blocks based on the pixel information in the identifier approximate pixel block and the pixel information in the target approximate pixel block to obtain the fine boundary of each target to be labeled. In the above embodiment, the traversal region is mainly obtained by obtaining the two coordinates of the cluster center and then determining the four boundaries based on the two coordinates and the preset distance. As for clustering approximate pixel clusters based on pixel information in the identified approximate pixel clusters and pixel information in the target approximate pixel clusters, this can be achieved by obtaining information such as the coordinates and color of the cluster center of the identified approximate pixel clusters and the coordinates and color of the cluster center of the target approximate pixel clusters; then, the distance on the plane and the color difference determined based on these coordinate information are used; finally, the target approximate pixel clusters are judged based on the distance on the plane and the color difference, so as to obtain the fine boundary of each target to be labeled based on the judgment result.
[0061] From the above Figure 1 As can be seen from the example image annotation method, unlike related technologies that suffer from high annotation costs and low accuracy due to manual annotation of image samples, the technical solution of this application first denoises the image to be annotated. The denoised image is used as the basis for calculation, avoiding the adverse effects of noise in the original image on the accuracy of subsequent calculations. On the other hand, after calculating the pixel gradient of the denoised image, secondary annotation is performed on each target to be annotated in the denoised image based on the pixel gradient of the denoised image. Since the secondary annotation is based on at least one manual annotation, automatic or semi-automatic annotation, it further adjusts the boundary towards the pixel direction of the gradient increase (i.e., the position where the color change is most drastic), making the boundary more accurate. Therefore, it not only has high annotation efficiency but also high annotation accuracy.
[0062] Corresponding to the aforementioned application function implementation method embodiments, this application also provides an image annotation device, an electronic device, and corresponding embodiments.
[0063] See Figure 4 This is a schematic diagram of the image annotation device shown in an embodiment of this application. For ease of explanation, only the parts related to the embodiments of this application are shown. Figure 4 The example image annotation device mainly includes a denoising module 401, a calculation module 402, and a secondary annotation module 403, wherein:
[0064] The denoising module 401 is used to denoise the image to be labeled to obtain a denoised image.
[0065] The calculation module 402 is used to calculate the pixel gradient of the denoised image;
[0066] The secondary annotation module 403 is used to perform secondary annotation on each target to be annotated in the denoised image based on the pixel gradient of the denoised image, so as to obtain the fine boundary of each target to be annotated.
[0067] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated further here.
[0068] From the above Figure 4 As can be seen from the example image annotation device, unlike related technologies that suffer from high annotation costs and low accuracy due to manual annotation when annotating image samples, the technical solution of this application first performs denoising processing on the image to be annotated. The resulting denoised image serves as the basis for calculation, avoiding the adverse effects of noise in the original image on the accuracy of subsequent calculations. On the other hand, after calculating the pixel gradient of the denoised image, secondary annotation is performed on each target to be annotated in the denoised image based on the pixel gradient of the denoised image. Since the secondary annotation is based on at least one manual, automatic or semi-automatic annotation, it further adjusts the boundary towards the pixel direction of the gradient increase (i.e., the position where the color change is most drastic), making the boundary more accurate. Therefore, it not only has high annotation efficiency but also high annotation accuracy.
[0069] Figure 5 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application.
[0070] See Figure 5 The electronic device 500 includes a memory 510 and a processor 520.
[0071] The processor 520 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.
[0072] Memory 510 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. ROM may store static data or instructions required by the processor 520 or other modules of the computer. Permanent storage devices may be read-write storage devices. Permanent storage devices may be non-volatile storage devices that retain stored instructions and data even when the computer is powered off. In some embodiments, permanent storage devices use mass storage devices (e.g., magnetic or optical disks, flash memory) as permanent storage devices. In other embodiments, permanent storage devices may be removable storage devices (e.g., floppy disks, optical drives). System memory may be a read-write storage device or a volatile read-write storage device, such as dynamic random access memory. System memory may store some or all of the instructions and data required by the processor during operation. Furthermore, memory 510 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (e.g., DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks may also be used. In some embodiments, memory 510 may include a removable storage device that is readable and / or writable, such as a laser disc (CD), a read-only digital multifunction optical disc (e.g., DVD-ROM, dual-layer DVD-ROM), a read-only Blu-ray disc, an ultra-high density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not contain carrier waves or transient electronic signals transmitted wirelessly or via wired connections.
[0073] The memory 510 stores executable code, which, when processed by the processor 520, can cause the processor 520 to execute part or all of the methods described above.
[0074] Furthermore, the method according to this application can also be implemented as a computer program or computer program product, which includes computer program code instructions for performing some or all of the steps in the method described above.
[0075] Alternatively, this application may be implemented as a computer-readable storage medium (or a non-transitory machine-readable storage medium or a machine-readable storage medium) storing executable code (or computer program or computer instruction code) thereon, which, when executed by a processor of an electronic device (or server, etc.), causes the processor to perform part or all of the steps of the methods described above according to this application.
[0076] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.
Claims
1. An image annotation method, characterized in that, The method includes: The image to be labeled is denoised to obtain a denoised image; Calculate the pixel gradient of the denoised image; Based on the pixel gradient of the denoised image, secondary annotation is performed on each target to be labeled in the denoised image to obtain the fine boundary of each target to be labeled; The step of performing secondary annotation on each target to be labeled in the denoised image based on the pixel gradient of the denoised image to obtain the fine boundary of each target to be labeled includes: Based on the pre-labeled data of the image to be labeled, the first boundary of each target to be labeled in the denoised image is obtained; wherein, the image to be labeled has been pre-labeled before the denoising process; the first boundary of each target to be labeled is moved in the direction of increasing pixel gradient until it stops at the maximum value of pixel gradient, and the second boundary of each target to be labeled obtained at the stop is taken as the fine boundary of each target to be labeled. Alternatively, based on the pixel gradient of the denoised image and the initial cluster centers determined by acquiring the first boundary of each target to be labeled, a second-generation cluster center is determined; the difference between the second-generation cluster center and each pixel in a preset neighborhood is calculated, and the pixels in the preset neighborhood are segmented according to the difference to form approximate pixel clusters; the approximate pixel clusters are clustered according to the pixel information corresponding to the region identifier defined by the first boundary and the pixel information in the approximate pixel clusters to obtain the fine boundary of each target to be labeled.
2. The image annotation method according to claim 1, characterized in that, The denoising process on the image to be labeled, to obtain a denoised image, includes: Determine a target mean square error image with the same size as the image to be labeled; Based on the preset blur radius and the target mean square error image, determine the target Gaussian blur kernel corresponding to each pixel in the image to be labeled; For each pixel in the image to be labeled, the target Gaussian blur value corresponding to the pixel is determined based on the target Gaussian blur kernel and the target pixels around the pixel that are within the preset blur radius, so as to obtain the Gaussian blurred image corresponding to the image to be labeled.
3. The image annotation method according to claim 1, characterized in that, The denoising process on the image to be labeled, to obtain a denoised image, includes: The image to be labeled and the corresponding reference image are subjected to mean filtering to construct a first noise characteristic map; and the image to be labeled and the reference image are subjected to downsampling and mean filtering in the target color channel to construct a second noise characteristic map. The noise intensity of the image to be labeled is determined by combining the first noise characteristic map and the brightness feature corresponding to the image to be labeled. Based on the noise intensity, a target noise characteristic map is determined by combining the first noise characteristic map and the second noise characteristic map, so as to remove noise from the image to be labeled according to the target noise characteristic map.
4. The image annotation method according to claim 1, characterized in that, The step of clustering the approximate pixel clusters based on the pixel information corresponding to the first boundary-defined region identifier and the pixel information in the approximate pixel clusters to obtain the fine boundaries of each target to be labeled includes: The target approximate pixel cluster is determined from multiple approximate pixel clusters, and the traversal area is determined based on the cluster center of the target approximate pixel cluster and a preset distance. Obtain the pixel clusters in the traversed region that approximate the identifier; Approximate pixel clusters are clustered based on the pixel information in the identified approximate pixel clusters and the pixel information in the target approximate pixel clusters to obtain the fine boundaries of each target to be labeled.
5. An image annotation device, characterized in that, The device includes: The denoising module is used to denoise the image to be labeled, so as to obtain a denoised image. A calculation module is used to calculate the pixel gradient of the denoised image; The secondary annotation module is used to perform secondary annotation on each target to be annotated in the denoised image based on the pixel gradient of the denoised image, so as to obtain the fine boundary of each target to be annotated. The step of performing secondary annotation on each target to be labeled in the denoised image based on the pixel gradient of the denoised image to obtain the fine boundary of each target to be labeled includes: Based on the pre-labeled data of the image to be labeled, the first boundary of each target to be labeled in the denoised image is obtained; wherein, the image to be labeled has been pre-labeled before the denoising process; the first boundary of each target to be labeled is moved in the direction of increasing pixel gradient until it stops at the maximum value of pixel gradient, and the second boundary of each target to be labeled obtained at the stop is taken as the fine boundary of each target to be labeled. Alternatively, based on the pixel gradient of the denoised image and the initial cluster centers determined by acquiring the first boundary of each target to be labeled, a second-generation cluster center is determined; the difference between the second-generation cluster center and each pixel in a preset neighborhood is calculated, and the pixels in the preset neighborhood are segmented according to the difference to form approximate pixel clusters; the approximate pixel clusters are clustered according to the pixel information corresponding to the region identifier defined by the first boundary and the pixel information in the approximate pixel clusters to obtain the fine boundary of each target to be labeled.
6. An electronic device, characterized in that, include: processor; as well as A memory having executable code stored thereon, which, when executed by the processor, causes the processor to perform the method as described in any one of claims 1 to 4.
7. A computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method as claimed in any one of claims 1 to 4.