A foreground object cutout method and system

Abstract

The present application discloses a method and system for foreground object clipping, comprising: inputting two images from different light sources into an illumination normalization module to obtain an illumination residual result; performing a tensor concatenation on the channel dimension based on the illumination residual result and the two downsampled images; obtaining a tensor based on a convolution calculation, the tensor containing the original information and the illumination residual result information; inputting the tensor into a Unified Network (UNet) network, which is used to extract the alpha value of the transparent channel of the foreground; performing a tensor concatenation on the result obtained by the UNet network and the illumination residual result; and performing a convolution calculation on the result of the tensor concatenation to obtain a clipping result image. This method aims to address the issues of inputting background and image models in deep learning methods, enabling a more efficient, automated process and higher accuracy for the task of extracting foreground objects.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of deep learning technology, and specifically to a method and system for cutting out a foreground object. Background Art

[0002] Traditional video processing requires extracting a foreground object to replace the background. In such scenarios, it's necessary to obtain the transparency channel value of the foreground object—that is, the transparency value of the foreground object at each pixel. Foreground objects come in a wide variety of types, and cutouts, hairy edges, and transparent areas are difficult to handle. Manual extraction is time-consuming and inaccurate due to pixel-level annotation. Summary of the Invention

[0003] To this end, the embodiments of the present application provide a foreground object cutout method and system, which aims to solve the problems existing in the input background and image models in deep learning methods, so that the task of obtaining foreground objects has a more efficient automated process and higher accuracy.

[0004] In order to achieve the above objectives, the embodiments of the present application provide the following technical solutions:

[0005] According to a first aspect of an embodiment of the present application, a method for cutting out a foreground object is provided, the method comprising:

[0006] Get two pictures from different light sources;

[0007] Input the illumination normalization module to obtain an illumination residual result;

[0008] Performing tensor concatenation on the channel dimension based on the illumination residual result and the two downsampled images;

[0009] A tensor is obtained based on convolution calculation, wherein the tensor contains original information and illumination residual result information;

[0010] Input the tensor into a UNet network, and the UNet network is used to extract the alpha value of the transparent channel of the foreground;

[0011] Perform a tensor connection between the result obtained by the UNet network and the illumination residual result;

[0012] Perform convolution calculation on the result of tensor connection to obtain the cutout result image.

[0013] Optionally, the input illumination normalization module obtains an illumination residual result, including:

[0014] Perform several convolution calculations on two images from different light sources to obtain two illumination normalized results;

[0015] The two illumination normalization results are subtracted based on the corresponding elements to obtain the illumination residual results of two pictures from different light sources.

[0016] Optionally, each convolution calculation performed on two pictures from different light sources is based on different convolution parameters, and the convolution parameters are based on different convolution models.

[0017] Optionally, the convolution calculation includes the following steps: performing a sliding frame operation on the input image with a k*k convolution kernel, each sliding frame corresponds to the k*k size area matrix of the image, and performing a dot product operation on the two k*k matrices to obtain a numerical result. After the sliding frame is completed, the multiple numerical results form a new matrix, that is, the new image output after the convolution operation, and the new image is used as the input of the next convolution kernel.

[0018] According to a second aspect of an embodiment of the present application, a foreground object cutout system is provided, the system comprising:

[0019] Input module, used to obtain two pictures from different light sources;

[0020] The illumination normalization module is used to obtain an illumination residual result;

[0021] A tensor connection module, configured to perform tensor connection on the channel dimension based on the illumination residual result and the two downsampled images;

[0022] A convolution module is used to obtain a tensor based on convolution calculation, wherein the tensor contains original information and illumination residual result information;

[0023] A UNet network module, used to input the tensor into a UNet network, and the UNet network is used to extract the alpha value of the transparent channel of the foreground;

[0024] The tensor connection module is further used to perform a tensor connection between the result obtained by the UNet network and the illumination residual result;

[0025] The convolution module is also used to perform convolution calculation on the result of tensor connection to obtain a cutout result image.

[0026] Optionally, the illumination normalization module is specifically configured to:

[0027] Perform several convolution calculations on two images from different light sources to obtain two illumination normalized results; subtract the two illumination normalized results based on corresponding elements to obtain the illumination residual results of the two images from different light sources.

[0028] Optionally, each convolution calculation performed on two pictures from different light sources is based on different convolution parameters, and the convolution parameters are based on different convolution models.

[0029] Optionally, the convolution calculation includes the following steps: performing a sliding frame operation on the input image with a k*k convolution kernel, each sliding frame corresponds to the k*k size area matrix of the image, and performing a dot product operation on the two k*k matrices to obtain a numerical result. After the sliding frame is completed, the multiple numerical results form a new matrix, that is, the new image output after the convolution operation, and the new image is used as the input of the next convolution kernel.

[0030] According to a third aspect of an embodiment of the present application, a device is provided, comprising: a data acquisition device, a processor and a memory; the data acquisition device is used to acquire data; the memory is used to store one or more program instructions; the processor is used to execute one or more program instructions to execute any method described in the first aspect.

[0031] According to a fourth aspect of an embodiment of the present application, a computer-readable storage medium is provided, wherein the computer storage medium contains one or more program instructions, and the one or more program instructions are used to execute the method as described in any one of the first aspects.

[0032] In summary, the embodiments of the present application provide a foreground object cutout method and system. Two images from different light sources are input into an illumination normalization module to obtain an illumination residual result. A tensor connection is performed on the channel dimension based on the illumination residual result and the two downsampled images. A tensor is obtained based on a convolution calculation, and the tensor contains the original information and the illumination residual result information. The tensor is input into a UNet network, which is used to extract the alpha value of the transparent channel of the foreground. The result obtained by the UNet network is tensor-connected with the illumination residual result. A convolution calculation is performed on the result of the tensor connection to obtain a cutout result image. This method aims to solve the problems existing in inputting background and image models in deep learning methods, so that the task of obtaining foreground objects has a more efficient automated process and higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.

[0034] The structures, proportions, sizes, etc. illustrated in this specification are intended solely to complement the contents disclosed herein and to facilitate understanding and reading by persons skilled in the art. They are not intended to limit the conditions under which the present invention may be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportions, or adjustments in sizes, without affecting the efficacy and objectives of the present invention, shall remain within the scope of the technical contents disclosed herein.

[0035] Figure 1 A schematic diagram of the process of the foreground object cutout method provided in an embodiment of the present application;

[0036] Figure 2 A schematic diagram of the overall structure of the model provided in the embodiment of the present application;

[0037] Figure 3 A schematic diagram of the workflow of the illumination normalization module provided in an embodiment of the present application;

[0038] Figure 4 A block diagram of a foreground object cutout system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0039] The following describes the implementation of the present invention using specific embodiments. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. Obviously, the embodiments described are only a portion of the present invention, not all of it. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0040] Image cutout technology can be divided into traditional methods and deep learning methods based on the degree of automation. Traditional methods require not only the original image as input but also a manually input trimap image, that is, a mask corresponding to the original image. This mask includes parts that are completely foreground, parts that are completely non-foreground, and parts that are uncertain whether they are foreground. Traditional methods use the trimap image to infer the transparency channel of the uncertain parts from the pixels in the confirmed parts. Different inference methods derive different algorithms.

[0041] Deep learning methods do not require a trimap image as input, but can be divided into two methods based on whether a background image is required. Deep learning methods that do not require a background image only need to input an image to obtain a foreground transparent channel, while neural networks that also input a background image can obtain more refined results. The background image does not need to be exactly the same as the background in the input image, and can have some offset in spatial position and some changes in lighting.

[0042] Traditional methods achieve better cropping results, but require the input of manually annotated trimaps, which means the entire process is still not fully automated. Deep learning methods automate the entire process, but models without background image input perform poorly, especially for irregularly shaped, non-rectangular objects, where it is difficult to define the edges of the objects. Deep learning methods that require background input can achieve relatively good results, but are currently mainly used in real-time video calls, so they mostly target people. For objects with ever-changing shapes, especially the problem of some of the aforementioned people not existing, it is still full of challenges. In some scenarios, obtaining a background image is a relatively difficult task, and the object needs to be moved away before it can be captured through shooting. This requires the object to be prepared in advance before the scene.

[0043] This embodiment of the application proposes a method for image cropping that takes two images as input. These images are images of the same scene, but with different lighting. "Different lighting" here refers to different light sources, including differences in position, number, intensity, and color. Because different light sources result in different reflection and refraction effects, the foreground and background in the real world have different color values ​​after imaging. This color difference provides valuable information for foreground cropping.

[0044] Since different light sources will cause different reflection and refraction effects, the foreground and background in the real physical world will have different degrees of color values ​​after imaging. Such color difference is effective information in the foreground cutout task. To put it bluntly, for example, for the same background and object, the first time the lights are turned on on the top, bottom, left and right sides of the object, then a fully lit object is obtained, but because the object itself has reflections, this is an interference, and the reflective part in the photo may be considered as the background; then when shooting the second time, the scene and object do not change, the camera position does not change, but only the backlight is turned on, then there will be no reflection, but the light is relatively dark, and the boundary between the object and the ground may not be clear. The two photos taken in this way have advantages and disadvantages, but if used together, the disadvantages complement each other and complete object shape information can be obtained.

[0045] Figure 1 A method for cutting out a foreground object provided by an embodiment of the present application is shown, the method comprising:

[0046] Step 101: Acquire two images from different light sources;

[0047] Step 102: Input the illumination normalization module to obtain an illumination residual result;

[0048] Step 103: Perform tensor connection on the channel dimension based on the illumination residual result and the two downsampled images;

[0049] Step 104: Obtain a tensor based on the convolution calculation, wherein the tensor includes the original information and the illumination residual result information;

[0050] Step 105: Input the tensor into an image segmentation UNet network, where the UNet network is used to extract the alpha value of the transparent channel of the foreground;

[0051] Step 106: Perform a tensor connection between the result obtained by the UNet network and the illumination residual result;

[0052] Step 107: Perform convolution calculation on the result of tensor connection to obtain the cutout result image.

[0053] In one possible implementation, in step 102, the input illumination normalization module obtains an illumination residual result, including: performing several convolution calculations on two images from different light sources to obtain two illumination normalization results; and subtracting the two illumination normalization results based on corresponding elements to obtain illumination residual results of the two images from different light sources.

[0054] In a possible implementation, each convolution calculation performed on two images from different light sources is based on different convolution parameters, and the convolution parameters are based on different convolution models.

[0055] In one possible implementation, the convolution calculation includes the following steps: performing a sliding frame operation on the input image with a k*k convolution kernel, each sliding frame corresponds to the k*k size area matrix of the image, and performing a dot product operation on the two k*k matrices to obtain a numerical result. After the sliding frame is completed, the multiple numerical results form a new matrix, that is, the new image output after the convolution operation, and the new image is used as the input of the next convolution kernel.

[0056] Figure 2 A schematic diagram of the overall structure of the model applicable to the method provided in the embodiment of the present application is shown. Since UNet is a general structure, the details are omitted.

[0057] Input two pictures with different lighting. The two pictures are first passed through the illumination normalization module to obtain a residual result. This result is concatenated with the downsampled pictures of the two pictures in the channel dimension. The convolution calculation obtains a tensor containing the original information and the illumination difference information. The tensor is input into the UNet network structure. The UNet network structure is a module commonly used to extract the transparent channel (alpha) value of the foreground. The result is then combined with the previous illumination residual (the "pixel difference map" is the "illumination residual", that is, Figure 3 Because the illumination residual has more information that is helpful for extracting foreground objects, this information is reused and convolution is performed to obtain the final result.

[0058] Concat is a tensor connection operation, that is, two tensors are spliced ​​together. Figure 2 In the figure, after downsampling, a tensor of h*w*c is obtained. At the same time, the illumination residual is also a tensor of h*w*c. They are spliced ​​in the dimension of c to obtain a tensor of h*w*(2c).

[0059] Tensor is the English word for tensor, which is a variable. Tensor1 represents a variable.

[0060] Convolution is a fundamental computational method in deep learning. It's a tensor-based computation that uses a sliding frame to perform matrix dot products with the same parameters across different locations in a tensor. Convolution, abbreviated as conv, is a fundamental operation in deep learning networks, similar to operators like addition and subtraction in basic mathematics.

[0061] It is a basic computing unit in the deep network. The specific operation is to use a k*k convolution kernel (essentially a k*k matrix) to perform a sliding operation on the input image. Each sliding operation corresponds to a k*k area of ​​the image (essentially also a k*k matrix). The two k*k matrices are multiplied to obtain a numerical result. After the sliding operation is completed, multiple numerical results form a new matrix, which is the new image output after the convolution operation. This new image is used as the input of the next convolution kernel.

[0062] The role of this "CONV" module in the deep network can be understood as the "multiplication" operator in the ordinary equation. They are both basic computing units in the calculation. The calculation operation is the same, but the parameters in each convolution kernel are different. This parameter is obtained during model training.

[0063] The model inputs two images taken under different lighting conditions, and uses these two images to determine the transparency channel (alpha) value of the foreground object. In the model's structural design, a twin network module for "lighting normalization" was designed to maintain the same foreground in images under different lighting conditions.

[0064] Figure 3 The following diagram shows the workflow of the illumination normalization module. The illumination normalization module adjusts the illumination of the input image, using a neural network to normalize the illumination intensity to a standard value, so that images under different lighting conditions are output with uniform illumination. This output modifies the overall image. Because the illumination normalization module does not learn the reflection and refraction effects of different objects, it lacks uniform sensitivity to the textures of different materials.

[0065] The illumination normalization module is a twin network. Two pictures with different illumination are input into the same network. The output results are then subtracted from each other to obtain the pixel difference map brought by different texture materials. The pixel difference map is the result of subtracting the corresponding elements and is the illumination residual. Figure 3 The right side of the flow chart.

[0066] For two tensors, the length and width of the two tensors are the same as the number of channels. Figure 3 Their shapes are identical: (w / 2)*(h / 2)*c. The number of channels, c, is not shown in the diagram. A specific position in one tensor corresponds to the same position in the other tensor; this is called a "corresponding element." Halving both the length and width is an empirical result based on extensive experimentation, but theoretically, it's not the only possible scenario. A new shape of any size can be created. However, the shapes of the "Normalized by Illumination 1" and "Normalized by Illumination 2" must be identical to allow for the next step of element-by-element subtraction.

[0067] For example, if the two input tensors are t1 and t2, both of shape (w / 2)*(h / 2)*c, then the value at position (1,1,5) in t1 "corresponds to the element in t2" is the value at position (1,1,5) in t2. Therefore, after "subtracting corresponding elements" from the two tensors, the result is a "pixel difference map" of (w / 2)*(h / 2)*3.

[0068] In summary, the embodiments of the present application provide a foreground object cutout method, which obtains an illumination residual result by inputting two images from different light sources into an illumination normalization module; performs a tensor connection on the channel dimension based on the illumination residual result and the two downsampled images; obtains a tensor based on convolution calculation, the tensor containing the original information and the illumination residual result information; inputs the tensor into a UNet network, which is used to extract the alpha value of the transparent channel of the foreground; performs a tensor connection on the result obtained by the UNet network and the illumination residual result; and performs a convolution calculation on the result of the tensor connection to obtain the cutout result image. The method aims to solve the problems existing in the input background and image models of deep learning methods, so that the task of obtaining foreground objects has a more efficient automated process and higher accuracy.

[0069] Based on the same technical concept, the embodiment of the present application also provides a foreground object cutout system, such as Figure 4 As shown, the system includes:

[0070] Input module 401, used to obtain two pictures from different light sources;

[0071] The illumination normalization module 402 is used to obtain an illumination residual result;

[0072] A tensor connection module 403 is configured to perform tensor connection in a channel dimension based on the illumination residual result and the two downsampled images;

[0073] A convolution module 404 is configured to obtain a tensor based on convolution calculation, wherein the tensor includes original information and illumination residual result information;

[0074] A UNet network module 405 is used to input the tensor into a UNet network, and the UNet network is used to extract the alpha value of the transparent channel of the foreground;

[0075] The tensor connection module 403 is further used to perform a tensor connection between the result obtained by the UNet network and the illumination residual result;

[0076] The convolution module 404 is further configured to perform convolution calculation on the result of tensor connection to obtain a cutout result image.

[0077] In a possible implementation, the illumination normalization module 402 is specifically configured to:

[0078] Perform several convolution calculations on two images from different light sources to obtain two illumination normalized results; subtract the two illumination normalized results based on corresponding elements to obtain the illumination residual results of the two images from different light sources.

[0079] In a possible implementation, each convolution calculation performed on two images from different light sources is based on different convolution parameters, and the convolution parameters are based on different convolution models.

[0080] In one possible implementation, the convolution calculation includes the following steps: performing a sliding frame operation on the input image with a k*k convolution kernel, each sliding frame corresponds to the k*k size area matrix of the image, and performing a dot product operation on the two k*k matrices to obtain a numerical result. After the sliding frame is completed, the multiple numerical results form a new matrix, that is, the new image output after the convolution operation, and the new image is used as the input of the next convolution kernel.

[0081] Based on the same technical concept, an embodiment of the present application also provides a device, which includes: a data acquisition device, a processor and a memory; the data acquisition device is used to collect data; the memory is used to store one or more program instructions; the processor is used to execute one or more program instructions to execute the described method.

[0082] Based on the same technical concept, an embodiment of the present application further provides a computer-readable storage medium, wherein the computer storage medium contains one or more program instructions, and the one or more program instructions are used to execute the described method.

[0083] In this specification, the various embodiments of the above method are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. For related parts, please refer to the partial description of the method embodiment.

[0084] It should be noted that although the operations of the method of the present invention are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in this specific order, or that all illustrated operations must be performed to achieve the desired results. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0085] Although the present application provides method operation steps such as embodiments or flow charts, more or fewer operation steps may be included based on conventional or non-creative means. The order of steps listed in the embodiments is only one way of executing the order of many steps and does not represent a unique execution order. When the device or client product in practice is executed, it can be executed in sequence or in parallel according to the method shown in the embodiments or the drawings (for example, a parallel processor or a multi-threaded processing environment, or even a distributed data processing environment). The term "comprise", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, product or equipment including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such process, method, product or equipment. In the absence of more restrictions, it is not excluded that there are other identical or equivalent elements in the process, method, product or equipment including the elements.

[0086] The units, devices or modules illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. For the convenience of description, the above devices are described in terms of functions and are divided into various modules and described separately. Of course, when implementing this application, the functions of each module can be implemented in the same or multiple software and / or hardware, or the modules that implement the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0087] Those skilled in the art will also appreciate that, in addition to implementing the controller in pure computer-readable program code, it is entirely possible to implement the same functionality by logically programming the method steps in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, and the like. Therefore, such a controller can be considered a hardware component, and the devices included therein for implementing various functions can also be considered structures within the hardware component. Alternatively, the devices for implementing various functions can be considered both software modules implementing the method and structures within the hardware component.

[0088] The present application may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc. that perform specific tasks or implement specific abstract data types. The present application may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communications network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0089] Through the description of the above implementation methods, it can be seen that those skilled in the art can clearly understand that the present application can be implemented by means of software plus the necessary general hardware platform. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a mobile terminal, a server, or a network device, etc.) to execute the methods described in various embodiments of the present application or certain parts of the embodiments.

[0090] The various embodiments in this specification are described in a progressive manner. References to the same or similar parts between the various embodiments are sufficient. Each embodiment focuses on the differences from the other embodiments. The present application can be used in a variety of general-purpose or specialized computer system environments or configurations. For example, personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments that include any of the above systems or devices.

[0091] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A foreground object cutout method, characterized in that: The method comprises: Get two pictures from different light sources; Input the illumination normalization module to obtain an illumination residual result. Specifically, the module performs several convolution operations on two images from different light sources to obtain two illumination normalization results. The module then subtracts the two illumination normalization results based on corresponding elements to obtain the illumination residual result of the two images from different light sources. Performing a tensor connection on the channel dimension based on the illumination residual result and the downsampled images of the two images from different light sources to obtain a first tensor; Performing a convolution operation on the first tensor to obtain a second tensor, where the second tensor includes original information and illumination residual result information. Input the second tensor into a UNet network, wherein the UNet network is used to extract the alpha value of the transparent channel of the foreground; Perform a tensor connection between the result obtained by the UNet network and the illumination residual result to obtain a third tensor; A convolution calculation is performed on the third tensor to obtain a cutout result image.

2. The method according to claim 1, wherein Each convolution calculation performed on two images from different light sources is based on different convolution parameters, which are based on different convolution models.

3. The method according to claim 2, wherein The convolution calculation includes the following steps: a k*k convolution kernel is used to perform a sliding operation on the input image. Each sliding operation corresponds to a k*k area matrix of the image. Two k*k matrices are subjected to a dot product operation to obtain a numerical result. After the sliding operation is completed, multiple numerical results form a new matrix, that is, a new image output after the convolution operation. The new image is used as the input of the next convolution kernel.

4. A foreground object cutout system, characterized in that: The system comprises: Input module, used to obtain two pictures from different light sources; The illumination normalization module is used to obtain an illumination residual result. Specifically, it includes: performing several convolution calculations on two images from different light sources to obtain two illumination normalization results; subtracting the two illumination normalization results based on corresponding elements to obtain the illumination residual result of the two images from different light sources; A tensor connection module is configured to perform a tensor connection on a channel dimension based on the illumination residual result and the downsampled images of the two images from different light sources to obtain a first tensor; A convolution module, configured to obtain a second tensor based on a convolution calculation of the first tensor, where the second tensor includes original information and illumination residual result information; A UNet network module, used to input the second tensor into a UNet network, wherein the UNet network is used to extract an alpha value of a transparent channel of a foreground; The tensor connection module is further used to perform a tensor connection on the result obtained by the UNet network and the illumination residual result to obtain a third tensor; The convolution module is further used to perform convolution calculation on the third tensor to obtain a cutout result image.

5. The system according to claim 4, wherein: Each convolution calculation performed on two images from different light sources is based on different convolution parameters, which are based on different convolution models.

6. The system according to claim 5, wherein: The convolution calculation includes the following steps: a k*k convolution kernel is used to perform a sliding operation on the input image. Each sliding operation corresponds to a k*k area matrix of the image. Two k*k matrices are subjected to a dot product operation to obtain a numerical result. After the sliding operation is completed, multiple numerical results form a new matrix, that is, a new image output after the convolution operation. The new image is used as the input of the next convolution kernel.

7. A device, characterized in that The device includes: a data acquisition device, a processor and a memory; The data acquisition device is used to acquire data; the memory is used to store one or more program instructions; and the processor is used to execute one or more program instructions to perform the method according to any one of claims 1 to 3.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium includes one or more program instructions, and the one or more program instructions are used to execute the method according to any one of claims 1 to 3.

Images