Laplacian pyramid image processing method and device
A Laplacian pyramid image processing method that performs downsampling followed by convolution and upsampling on an FPGA chip solves the problem of high computational resource consumption in traditional methods and achieves more efficient image processing.
Patent Information
- Application Number
- CN202510754722.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-06
- Publication Date
- 2025-10-31
AI Technical Summary
Traditional Laplacian pyramid image processing schemes consume a lot of computational resources and are inefficient, especially in high-resolution images and deep neural networks where the computational load increases exponentially.
An FPGA chip is used to execute the Laplacian pyramid image processing method. By downsampling followed by convolution and upsampling, the computational complexity is reduced. This includes convolutional fusion of downsampled images, convolutional fusion of upsampled images, and subtraction of the final image.
This reduces computational complexity and chip computing resource consumption, while improving processing efficiency.
Smart Images
Figure CN120876243A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of image processing technology, specifically relating to a method and apparatus for processing Laplacian pyramid images. Background Technology
[0002] The Laplacian pyramid is a commonly used multi-resolution image processing method. It constructs image pyramids of different resolutions by repeatedly performing Gaussian convolution and downsampling operations on the image. Then, Laplacian difference is applied to the pyramids to achieve image decomposition and reconstruction. Specifically, the Laplacian pyramid consists of a series of Gaussian pyramids. Each level of the Gaussian pyramid is obtained by sequentially performing Gaussian blur and downsampling on the original image, and the Laplacian pyramid is obtained by subtracting two adjacent levels of the Gaussian pyramid.
[0003] In implementing the Laplacian pyramid, each pyramid layer requires a convolution calculation. Traditional convolution operations, however, require a complete convolution calculation for each pixel. Furthermore, to prevent feature loss, traditional Laplacian pyramid implementations typically employ a sequence of convolution followed by downsampling, then upsampling followed by convolution, and finally image subtraction. Figure 1 As shown. This means that each convolutional kernel requires a significant amount of computational resources, resulting in a computationally intensive and inefficient implementation. Furthermore, in traditional convolutional operations, the larger the kernel size and the number of channels, the more computational resources are required. For high-resolution images and deep neural networks, this can lead to an exponential increase in computational load. Summary of the Invention
[0004] The purpose of this application is to propose a method and apparatus for Laplacian pyramid image processing, so as to solve the technical problems of high computational resource consumption and low efficiency in existing Laplacian pyramid-based image processing schemes.
[0005] To address the aforementioned technical problems, this application provides a method for processing Laplacian pyramid images, employing the following technical solution:
[0006] A method for Laplacian pyramid image processing, executed by an FPGA chip, includes:
[0007] Receive an input image, perform a first downsampling on the input image to obtain a first downsampling image and a second downsampling image, perform a first convolution on the first downsampling image to obtain a first convolutional image, and perform a second convolution on the second downsampling image to obtain a second convolutional image, and perform image fusion on the first convolutional image and the second convolutional image to obtain a first fused image;
[0008] The first fused image is downsampled a second time to obtain a third downsampled image and a fourth downsampled image. The third downsampled image is convolved to obtain a third convolved image, and the fourth downsampled image is convolved to obtain a fourth convolved image. The third convolved image and the fourth convolved image are fused to obtain a second fused image.
[0009] The second fused image is subjected to a third convolution to obtain a fifth convolution image and a sixth convolution image. The fifth convolution image is subjected to a first upsampling to obtain a first upsampled image, and the sixth convolution image is subjected to a first upsampling to obtain a second upsampled image. The first upsampled image and the second upsampled image are fused to obtain a third fused image.
[0010] The third fused image is convolved a fourth time to obtain the seventh and eighth convolved images. The seventh convolved image is upsampled a second time to obtain the third upsampled image, and the eighth convolved image is upsampled a second time to obtain the fourth upsampled image. The third upsampled image and the fourth upsampled image are fused to obtain the fourth fused image.
[0011] The input image and the fourth fused image are subtracted to obtain the difference image, which is then used as the output image of the Laplacian pyramid.
[0012] To address the aforementioned technical problems, this application also provides an apparatus for processing Laplacian pyramid images, employing the following technical solution:
[0013] An apparatus for Laplacian pyramid image processing, wherein the Laplacian pyramid image processing apparatus is disposed in an FPGA chip, the apparatus comprising:
[0014] The first downsampling processing module is used to receive an input image, perform a first downsampling on the input image to obtain a first downsampling image and a second downsampling image, perform a first convolution on the first downsampling image to obtain a first convolutional image, perform a second convolution on the second downsampling image to obtain a second convolutional image, and perform image fusion on the first convolutional image and the second convolutional image to obtain a first fused image.
[0015] The second downsampling processing module is used to perform a second downsampling on the first fused image to obtain a third downsampling image and a fourth downsampling image, to perform convolution on the third downsampling image to obtain a third convolutional image, and to perform convolution on the fourth downsampling image to obtain a fourth convolutional image, and to perform image fusion on the third convolutional image and the fourth convolutional image to obtain a second fused image.
[0016] The first upsampling processing module is used to perform a third convolution on the second fused image to obtain a fifth convolution image and a sixth convolution image, perform a first upsampling on the fifth convolution image to obtain a first upsampled image, and perform a first upsampling on the sixth convolution image to obtain a second upsampled image, and perform image fusion on the first upsampled image and the second upsampled image to obtain a third fused image.
[0017] The second upsampling processing module is used to perform a fourth convolution on the third fused image to obtain a seventh convolution image and an eighth convolution image, perform a second upsampling on the seventh convolution image to obtain a third upsampled image, and perform a second upsampling on the eighth convolution image to obtain a fourth upsampled image, and perform image fusion on the third upsampled image and the fourth upsampled image to obtain a fourth fused image.
[0018] The image subtraction module is used to subtract the input image from the fourth fused image to obtain the difference image, and then use the difference image as the output image of the Laplacian pyramid.
[0019] To address the aforementioned technical problems, this application also provides a computer device that employs the following technical solution:
[0020] A computer device includes a memory and a processor, the memory storing computer-readable instructions, the processor executing the computer-readable instructions to implement the steps of the Laplacian pyramid image processing method as described in any of the preceding claims.
[0021] To address the aforementioned technical problems, this application also provides a computer-readable storage medium, employing the technical solution described below:
[0022] A computer-readable storage medium storing computer-readable instructions that, when executed by a processor, implement the steps of the method for Laplacian pyramid image processing as described above.
[0023] Compared with the prior art, the embodiments of this application have the following main advantages:
[0024] This application discloses a method and apparatus for Laplacian pyramid image processing, belonging to the field of image processing technology. This application involves downsampling an input image to obtain two downsampled images, then convolving the two downsampled images and fusing the resulting convolutional images; downsampling the fused image again to obtain two new downsampled images, convolving the two new downsampled images, and fusing the resulting convolutional images again; convolving the fused image with two convolution kernels to obtain two convolutional images, then upsampling the two convolutional images to obtain two upsampled images, fusing the two upsampled images; convolving the new fused images with two convolution kernels to obtain two new convolutional images, upsampling the two new convolutional images to obtain two new upsampled images, and fusing the two new upsampled images; finally, subtracting the input image from the final fused image yields a difference image, which is the output image of a certain layer in the Laplacian pyramid. Compared with the traditional Laplacian pyramid image processing method, this application adopts the method of downsampling before convolution and convolution before upsampling. By using two downsampling and two upsampling, the computational complexity is greatly reduced, and chip computing resources and energy consumption are reduced while generating the same result. Attached Figure Description
[0025] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 This diagram illustrates the structural block diagram of the traditional Laplacian pyramid image implementation process;
[0027] Figure 2 A flowchart of one embodiment of the Laplacian pyramid image processing method according to this application is shown;
[0028] Figure 3 A structural block diagram of a Laplacian pyramid image implementation process according to an embodiment of this application is shown;
[0029] Figure 4 A schematic diagram of one embodiment of the Laplacian pyramid image processing apparatus according to this application is shown. Detailed Implementation
[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.
[0031] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.
[0032] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.
[0033] In the embodiments of this application, the Laplacian pyramid image processing method disclosed in this application is mainly executed by an FPGA chip, and the Laplacian pyramid image processing device is mainly disposed in the FPGA chip.
[0034] FPGA stands for Field-Programmable Gate Array. It is a programmable logic device whose internal hardware circuitry can be reconfigured as needed. FPGA chips can be reprogrammed by programmers to perform specific tasks or complete specific functions. Compared to ASIC chips, FPGA chips have greater flexibility and programmability, and are therefore widely used in applications that require frequent changes to circuit functions, such as video and image processing, communications, industrial control, aerospace, and medical devices.
[0035] FPGA chips can be programmed using hardware description languages (such as VHDL and Verilog) to translate execution logic into a description of digital logic circuits, which is then downloaded into the FPGA chip. FPGA chips typically consist of programmable logic units, programmable registers, internal memory, and input / output pins. FPGA chips can also be integrated with other devices (such as processors, memory, and sensors) to form a complete system.
[0036] In the embodiments of this application, the FPGA chip possesses high parallel computing capabilities and flexible programmability, which can effectively accelerate the execution speed of the Laplacian pyramid image processing method described in this application. By converting the Laplacian pyramid image processing method of this application into a hardware description language (such as VHDL or Verilog) and implementing the algorithm on the FPGA chip, the processing speed and efficiency can be greatly improved. Furthermore, the FPGA chip can also be integrated with other processors and devices to implement more complex image processing tasks.
[0037] It should be noted that the Laplacian pyramid image processing method provided in this application embodiment can also be executed by a server, and correspondingly, the Laplacian pyramid image processing apparatus can also be set in the server.
[0038] Continue to refer to Figure 2 The diagram illustrates a flowchart of an embodiment of the Laplacian pyramid image processing method according to this application. Embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that utilize digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.
[0039] Foundational artificial intelligence technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.
[0040] The Laplacian pyramid image processing method disclosed in this application includes the following steps:
[0041] S201, receive an input image, perform a first downsampling on the input image to obtain a first downsampling image and a second downsampling image, perform a first convolution on the first downsampling image to obtain a first convolutional image, and perform a second convolution on the second downsampling image to obtain a second convolutional image, and perform image fusion on the first convolutional image and the second convolutional image to obtain a first fused image.
[0042] In this embodiment, the first downsampling is performed by downsampling according to the image pixel columns. After receiving the input image, the server performs the first downsampling on the input image according to the image pixel columns to obtain a first downsampled image and a second downsampled image. The first downsampled image is obtained by downsampling the odd-numbered columns of the input image, and the second downsampled image is obtained by downsampling the even-numbered columns of the input image. A convolutional kernel matching the first downsampled image is obtained, i.e., the first convolutional kernel. The first downsampled image is convolved using the first convolutional kernel to obtain a first convolutional image. A convolutional kernel matching the second downsampled image is obtained, i.e., the second convolutional kernel. The second downsampled image is convolved using the second convolutional kernel to obtain a second convolutional image. Then, the first convolutional image and the second convolutional image are fused to obtain a first fused image. Image fusion refers to adding the pixels of the corresponding pixels of the two images.
[0043] In one specific embodiment of this application, such as Figure 3 As shown, Figure 3 The input image of the Nth layer of the middle Laplacian pyramid is downsampled according to the image pixel column to obtain a first downsampled image x1 and a second downsampled image x2. Then, the first downsampled image x1 is convolved by the first convolutional kernel Hx1, and the second downsampled image x2 is convolved by the second convolutional kernel Hx2. The first convolutional image and the second convolutional image are added together to obtain a first fused image.
[0044] Further, the system receives an input image, performs a first downsampling on the input image to obtain a first downsampling image and a second downsampling image, performs a first convolution on the first downsampling image to obtain a first convolutional image, and performs a second convolution on the second downsampling image to obtain a second convolutional image, and fuses the first convolutional image and the second convolutional image to obtain a first fused image, specifically including:
[0045] Receive the input image and downsample the input image according to the preset first downsampling method to obtain the first downsampled image and the second downsampled image;
[0046] Obtain a first convolutional kernel that matches the first downsampled image, and obtain a second convolutional kernel that matches the second downsampled image;
[0047] The first downsampled image is convolved using the first convolutional kernel to obtain the first convolutional image, and the second downsampled image is convolved using the second convolutional kernel to obtain the second convolutional image.
[0048] The first convolutional image and the second convolutional image are fused to obtain the first fused image.
[0049] Furthermore, the first downsampling method involves downsampling according to image pixel columns. The input image is received and downsampled according to a preset first downsampling method to obtain a first downsampled image and a second downsampled image. Specifically, this includes:
[0050] Receive the input image, downsample the odd-numbered columns of the input image to obtain the first downsampled image;
[0051] The even-numbered columns of the input image are downsampled to obtain the second downsampled image.
[0052] S202, perform a second downsampling on the first fused image to obtain a third downsampling image and a fourth downsampling image, convolve the third downsampling image to obtain a third convolutional image, and convolve the fourth downsampling image to obtain a fourth convolutional image, and fuse the third convolutional image and the fourth convolutional image to obtain a second fused image.
[0053] In this embodiment, the second downsampling is performed by downsampling the image pixels row by row. The first fused image is then downsampled to obtain a third downsampled image and a fourth downsampled image. The third downsampled image is obtained by downsampling the odd-numbered rows of the first fused image, and the fourth downsampled image is obtained by downsampling the even-numbered rows of the first fused image. A convolutional kernel matching the third downsampled image is obtained, and the third downsampled image is convolved using this kernel to obtain a third convolutional image. Similarly, a convolutional kernel matching the fourth downsampled image is obtained, and the fourth downsampled image is convolved using this kernel to obtain a fourth convolutional image. Finally, the third and fourth convolutional images are fused to obtain a second fused image.
[0054] In the specific embodiment described above, the first fused image is downsampled according to the pixel rows to obtain a third downsampled image y1 and a fourth downsampled image y2. Then, the third downsampled image y1 is convolved by the third convolutional kernel Hy1, and the fourth downsampled image y2 is convolved by the fourth convolutional kernel Hy2. The resulting third and fourth convolutional images are then added together to obtain the second fused image, which is also used as the input image for the N+1 layer of the Laplacian pyramid.
[0055] Further, the first fused image is downsampled a second time to obtain a third downsampled image and a fourth downsampled image. The third downsampled image is convolved to obtain a third convolved image, and the fourth downsampled image is convolved to obtain a fourth convolved image. The third and fourth convolved images are then fused to obtain a second fused image, specifically including:
[0056] The first fused image is downsampled according to the preset second downsampling method to obtain the third and fourth downsampled images;
[0057] Obtain the third convolutional kernel that matches the third downsampled image, and obtain the fourth convolutional kernel that matches the fourth downsampled image;
[0058] The third downsampled image is convolved using the third convolutional sub-kernel to obtain the third convolutional image, and the fourth downsampled image is convolved using the fourth convolutional sub-kernel to obtain the fourth convolutional image.
[0059] The third and fourth convolutional images are fused to obtain the second fused image.
[0060] Furthermore, the second downsampling method involves downsampling according to image pixel rows. The first fused image is downsampled according to the preset second downsampling method to obtain a third downsampled image and a fourth downsampled image, specifically including:
[0061] The odd-numbered rows of the first fused image are downsampled to obtain the third downsampled image;
[0062] The even-numbered rows of the first fused image are downsampled to obtain the fourth downsampled image.
[0063] S203, perform a third convolution on the second fused image to obtain a fifth convolution image and a sixth convolution image, perform a first upsampling on the fifth convolution image to obtain a first upsampled image, and perform a first upsampling on the sixth convolution image to obtain a second upsampled image, and perform image fusion on the first upsampled image and the second upsampled image to obtain a third fused image.
[0064] In this embodiment, a fifth convolutional sub-kernel and a sixth convolutional sub-kernel are obtained. The fifth convolutional sub-kernel is used to convolve the second fused image to obtain a fifth convolutional image. The sixth convolutional sub-kernel is used to convolve the second fused image to obtain a sixth convolutional image. The fifth convolutional image is upsampled in even-numbered columns to obtain a first upsampled image. The sixth convolutional image is upsampled in odd-numbered columns to obtain a second upsampled image. The first upsampling method is upsampling according to the image pixel columns.
[0065] In the specific embodiment described above, the second fused image is convolved using the fifth convolutional sub-kernel Gx1 and the sixth convolutional sub-kernel Gx2 respectively to obtain the fifth convolutional image and the sixth convolutional image. Then, the even-numbered columns of the fifth convolutional image are upsampled to obtain the first upsampled image x3, and the odd-numbered columns of the sixth convolutional image are upsampled to obtain the second upsampled image x4. The first upsampled image x3 and the second upsampled image x4 are added together to obtain the third fused image.
[0066] Further, a third convolution is performed on the second fused image to obtain a fifth convolution image and a sixth convolution image. The fifth convolution image is then upsampled to obtain a first upsampled image, and the sixth convolution image is also upsampled to obtain a second upsampled image. Finally, the first upsampled image and the second upsampled image are fused to obtain a third fused image, specifically including:
[0067] Obtain the fifth and sixth convolutional sub-kernels, and use the fifth convolutional sub-kernel to convolve the second fused image to obtain the fifth convolutional image, and use the sixth convolutional sub-kernel to convolve the second fused image to obtain the sixth convolutional image;
[0068] The fifth convolutional image is upsampled according to the first upsampling method to obtain the first upsampled image, and the sixth convolutional image is upsampled according to the first upsampling method to obtain the second upsampled image;
[0069] The first upsampled image and the second upsampled image are fused to obtain the third fused image.
[0070] Furthermore, the first upsampling method involves upsampling according to image pixel columns. The fifth convolutional image is upsampled according to the first upsampling method to obtain the first upsampled image, and the sixth convolutional image is upsampled according to the first upsampling method to obtain the second upsampled image. Specifically, this includes:
[0071] Upsample the even-numbered columns of the fifth convolutional image to obtain the first upsampled image;
[0072] Upsample the odd-numbered columns of the sixth convolutional image to obtain the second upsampled image.
[0073] S204, perform a fourth convolution on the third fused image to obtain a seventh convolution image and an eighth convolution image, perform a second upsampling on the seventh convolution image to obtain a third upsampled image, and perform a second upsampling on the eighth convolution image to obtain a fourth upsampled image, and perform image fusion on the third upsampled image and the fourth upsampled image to obtain a fourth fused image.
[0074] In this embodiment, a seventh convolutional kernel and an eighth convolutional kernel are obtained. The seventh convolutional kernel is used to convolve the third fused image to obtain a seventh convolutional image. The eighth convolutional kernel is used to convolve the third fused image to obtain an eighth convolutional image. The seventh convolutional image is upsampled with even-numbered rows to obtain a third upsampled image. The eighth convolutional image is upsampled with odd-numbered rows to obtain a fourth upsampled image. The second upsampling method is upsampling according to the image pixel rows.
[0075] In the specific embodiment described above, the third fused image is convolved using the seventh convolutional sub-kernel Gy1 and the eighth convolutional sub-kernel Gy2 respectively to obtain the seventh convolutional image and the eighth convolutional image. Then, the even-numbered rows of the seventh convolutional image are upsampled to obtain the third upsampled image y3, and the odd-numbered rows of the eighth convolutional image are upsampled to obtain the fourth upsampled image y4. The third upsampled image y3 and the fourth upsampled image y4 are added together to obtain the fourth fused image.
[0076] Furthermore, a fourth convolution is performed on the third fused image to obtain a seventh and eighth convolutional image. A second upsampling is performed on the seventh convolutional image to obtain a third upsampled image, and a second upsampling is performed on the eighth convolutional image to obtain a fourth upsampled image. Finally, the third and fourth upsampled images are fused to obtain a fourth fused image, specifically including:
[0077] Obtain the seventh and eighth convolutional sub-kernels, and use the seventh convolutional sub-kernel to convolve the third fused image to obtain the seventh convolutional image, and use the eighth convolutional sub-kernel to convolve the third fused image to obtain the eighth convolutional image;
[0078] The seventh convolutional image is upsampled according to the second upsampling method to obtain the third upsampled image, and the eighth convolutional image is upsampled according to the second upsampling method to obtain the fourth upsampled image;
[0079] The third upsampled image and the fourth upsampled image are fused to obtain the fourth fused image.
[0080] Furthermore, the second upsampling method involves upsampling according to the image pixel rows. The seventh convolutional image is upsampled using the second upsampling method to obtain the third upsampled image, and the eighth convolutional image is upsampled using the second upsampling method to obtain the fourth upsampled image. Specifically, this includes:
[0081] Upsample the even-numbered rows of the seventh convolutional image to obtain the third upsampled image;
[0082] Upsample the odd-numbered rows of the eighth convolutional image to obtain the fourth upsampled image.
[0083] S205, perform a subtraction operation on the input image and the fourth fused image to obtain the difference image, and use the difference image as the output image of the Laplacian pyramid.
[0084] In this embodiment, the input image and the fourth fused image are subtracted to obtain a difference image, which is then used as the output image of the Laplacian pyramid.
[0085] In the specific embodiment described above, a subtraction operation is performed on the input image of the Nth layer of the Laplacian pyramid and the fourth fused image to obtain a difference image, which is then used as the output image of the Nth layer of the Laplacian pyramid. When implementing the Laplacian pyramid, steps S201 to S205 can be performed once for each layer of the Laplacian pyramid to obtain the output image of each layer, ultimately completing the construction of the Laplacian pyramid.
[0086] Furthermore, before receiving the input image and downsampling it according to a preset first downsampling method to obtain the first downsampled image and the second downsampled image, the process also includes:
[0087] Obtain the preset first and second separable convolution kernels;
[0088] The first separable convolution kernel is split to obtain a first convolution kernel and a second convolution kernel, wherein both the first convolution kernel and the second convolution kernel are one-dimensional convolution kernels;
[0089] The second separable convolution kernel is split to obtain a third convolution kernel and a fourth convolution kernel, both of which are one-dimensional convolution kernels.
[0090] The first, second, third, fourth, fifth, sixth, seventh, and eighth convolutional sub-kernels can be obtained by downsampling the one-dimensional convolutional kernel in a special way. For example, the first, second, fifth, and sixth convolutional sub-kernels can be obtained by downsampling the one-dimensional convolutional kernel using the first downsampling method, and the third, fourth, seventh, and eighth convolutional sub-kernels can be obtained by downsampling the one-dimensional convolutional kernel using the second downsampling method.
[0091] In this embodiment, the server obtains a preset first separable convolutional kernel and a preset second separable convolutional kernel. The first and second separable convolutional kernels are two-dimensional convolutional kernels of the same size, such as a 3x3 kernel or a 5x5 kernel, etc. The first and second separable convolutional kernels can be identical or two different kernels, and can be set according to image processing requirements. The first separable convolutional kernel is split to obtain a first convolutional kernel and a second convolutional kernel. The second separable convolutional kernel is then split to obtain a third convolutional kernel and a fourth convolutional kernel. All four kernels are one-dimensional convolutional kernels.
[0092] It should be noted that, in specific embodiments of this application, separable convolutional kernels can be preset. By splitting two separable convolutional kernels, four one-dimensional convolutional kernels are obtained. Each of these four one-dimensional convolutional kernels is downsampled to obtain eight convolutional sub-kernels and two separable convolutional kernels. Alternatively, four one-dimensional convolutional kernels can be preset directly, and each of these four one-dimensional convolutional kernels is downsampled to obtain eight convolutional sub-kernels.
[0093] Furthermore, after splitting the first separable convolutional kernel to obtain the first convolutional kernel and the second convolutional kernel, the process also includes:
[0094] The first convolutional kernel is downsampled using a first downsampling method to obtain a first convolutional sub-kernel and a second convolutional sub-kernel; the second convolutional kernel is downsampled using a second downsampling method to obtain a third convolutional sub-kernel and a fourth convolutional sub-kernel.
[0095] Furthermore, after splitting the second separable convolutional kernel to obtain the third and fourth convolutional kernels, the process also includes:
[0096] The third convolutional kernel is downsampled using the first downsampling method to obtain the fifth and sixth convolutional sub-kernels; the fourth convolutional kernel is downsampled using the second downsampling method to obtain the seventh and eighth convolutional sub-kernels.
[0097] Furthermore, the first convolutional kernel is downsampled using a first downsampling method to obtain a first convolutional sub-kernel and a second convolutional sub-kernel, specifically including:
[0098] The odd-numbered columns of the first convolutional kernel are downsampled to obtain the first convolutional sub-kernel; the even-numbered columns of the first convolutional kernel are downsampled to obtain the second convolutional sub-kernel.
[0099] Furthermore, the third convolutional kernel is downsampled using the first downsampling method to obtain the fifth and sixth convolutional sub-kernels, specifically including:
[0100] The odd-numbered columns of the third convolutional kernel are downsampled to obtain the fifth convolutional sub-kernel; the even-numbered columns of the third convolutional kernel are downsampled to obtain the sixth convolutional sub-kernel.
[0101] Furthermore, the second convolutional kernel is downsampled using a second downsampling method to obtain the third and fourth convolutional sub-kernels, specifically including:
[0102] The odd-numbered rows of the second convolutional kernel are downsampled to obtain the third convolutional sub-kernel; the even-numbered rows of the second convolutional kernel are downsampled to obtain the fourth convolutional sub-kernel.
[0103] Furthermore, the fourth convolutional kernel is downsampled using a second downsampling method to obtain the seventh and eighth convolutional sub-kernels, specifically including:
[0104] Downsampling the odd-numbered rows of the fourth convolutional kernel yields the seventh convolutional sub-kernel; downsampling the even-numbered rows of the fourth convolutional kernel yields the eighth convolutional sub-kernel.
[0105] In this embodiment, the Laplacian pyramid image processing method operates on an electronic device (e.g., Figure 1 The server shown can receive various commands, images, or data via wired or wireless connections. It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G / 5G connections, WiFi connections, Bluetooth connections, WiMAX connections, Zigbee connections, UWB (ultra-wideband) connections, and other currently known or future-developed wireless connection methods.
[0106] In one specific embodiment of this application, the first convolutional kernel is [1,2,3,2,1], and the second convolutional kernel is... Downsampling the odd-numbered columns of the first convolutional kernel yields the first convolutional sub-kernel [1,3,1]. Downsampling the even-numbered columns of the first convolutional kernel yields the second convolutional sub-kernel [2,2,0]. It should be noted that the first and second convolutional sub-kernels must have the same size during downsampling. Therefore, when downsampling the even-numbered columns of the first convolutional kernel, to ensure that the generated convolutional sub-kernels all have a size of 1*3, "0"s need to be padded to the end of the even-numbered column downsampling results, resulting in the second convolutional kernel [2,2,0].
[0107] In one specific embodiment of this application, the odd-numbered rows of the second convolutional kernel are downsampled to obtain the third convolutional sub-kernel. Downsampling the even-numbered rows of the second convolutional kernel yields the fourth convolutional sub-kernel. It should be noted that during downsampling, the third and fourth convolutional sub-kernels must have the same size. Therefore, when downsampling the second convolutional kernel, to ensure that the generated convolutional sub-kernels all have a size of 3*1, "0"s need to be padded to the end of the downsampling results in even-numbered rows to obtain the fourth convolutional kernel.
[0108] In one specific embodiment of this application, the third convolutional kernel is [2,3,4,3,2], and the fourth convolutional kernel is... Downsampling the odd-numbered columns of the third convolutional kernel yields the fifth convolutional sub-kernel [2,4,2]; downsampling the even-numbered columns of the third convolutional kernel yields the sixth convolutional sub-kernel [0,4,4]. It should be noted that the fifth and sixth convolutional sub-kernels must have the same size during downsampling. Therefore, when downsampling the even-numbered columns of the third convolutional kernel, to ensure that the generated convolutional sub-kernels all have a size of 1*3, "0"s need to be added before the even-numbered column downsampling results, resulting in the sixth convolutional sub-kernel [0,3,3].
[0109] In one specific embodiment of this application, the odd-numbered rows of the fourth convolutional kernel are downsampled to obtain the seventh convolutional sub-kernel. Downsampling the even-numbered rows of the fourth convolutional kernel yields the eighth convolutional sub-kernel. It should be noted that during downsampling, the seventh and eighth convolutional sub-kernels must have the same size. Therefore, when downsampling the fourth convolutional kernel, to ensure that the generated convolutional sub-kernels all have a size of 3*1, "0"s need to be padded to the beginning of the downsampling results in even-numbered rows to obtain the eighth convolutional kernel.
[0110] In the above embodiments, this application discloses a method for Laplacian pyramid image processing, belonging to the field of image processing technology. This application involves downsampling the input image according to a preset first method, then convolving the downsampled image and fusing the resulting convolutional images; downsampling the fused image according to a preset second method, convolving the new downsampled image, and fusing the resulting new convolutional images; convolving the generated fused image, upsampling the convolutional image according to the preset first method, and fusing the upsampled image; continuing to convolve the generated fused image, upsampling the convolutional image according to the preset second method, and fusing the upsampled image; and finally, subtracting the input image from the final fused image to obtain a difference image. This application significantly reduces computational complexity, reducing chip computing resources and energy consumption while generating the same result.
[0111] 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 instructing related hardware with computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium, and when executed, they can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).
[0112] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures 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, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.
[0113] Further reference Figure 4 As a response to the above Figure 2 The implementation of the method shown in this application provides an embodiment of a Laplacian pyramid image processing apparatus, which is similar to... Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to an FPGA chip.
[0114] like Figure 4 As shown, the Laplacian pyramid image processing apparatus 400 described in this embodiment is disposed in an FPGA chip, and the apparatus includes:
[0115] The first downsampling processing module 401 is used to receive an input image, perform a first downsampling on the input image to obtain a first downsampling image and a second downsampling image, perform a first convolution on the first downsampling image to obtain a first convolutional image, perform a second convolution on the second downsampling image to obtain a second convolutional image, and perform image fusion on the first convolutional image and the second convolutional image to obtain a first fused image.
[0116] The second downsampling processing module 402 is used to perform a second downsampling on the first fused image to obtain a third downsampling image and a fourth downsampling image, to perform convolution on the third downsampling image to obtain a third convolutional image, and to perform convolution on the fourth downsampling image to obtain a fourth convolutional image, and to perform image fusion on the third convolutional image and the fourth convolutional image to obtain a second fused image.
[0117] The first upsampling processing module 403 is used to perform a third convolution on the second fused image to obtain a fifth convolution image and a sixth convolution image, perform a first upsampling on the fifth convolution image to obtain a first upsampled image, and perform a first upsampling on the sixth convolution image to obtain a second upsampled image, and perform image fusion on the first upsampled image and the second upsampled image to obtain a third fused image.
[0118] The second upsampling processing module 404 is used to perform a fourth convolution on the third fused image to obtain a seventh convolution image and an eighth convolution image, perform a second upsampling on the seventh convolution image to obtain a third upsampled image, and perform a second upsampling on the eighth convolution image to obtain a fourth upsampled image, and perform image fusion on the third upsampled image and the fourth upsampled image to obtain a fourth fused image.
[0119] The image subtraction module 405 is used to perform a subtraction operation on the input image and the fourth fused image to obtain a difference image, and uses the difference image as the output image of the Laplacian pyramid.
[0120] Furthermore, the first downsampling processing module 401 specifically includes:
[0121] The first downsampling submodule is used to receive the input image and downsample the input image according to the preset first downsampling method to obtain the first downsampled image and the second downsampled image.
[0122] The first sub-kernel acquisition submodule is used to acquire the first convolutional sub-kernel that matches the first downsampled image, and to acquire the second convolutional sub-kernel that matches the second downsampled image;
[0123] The first image convolution submodule is used to convolve the first downsampled image using the first convolution subkernel to obtain the first convolution image, and to convolve the second downsampled image using the second convolution subkernel to obtain the second convolution image.
[0124] The first image fusion submodule is used to fuse the first convolutional image and the second convolutional image to obtain the first fused image;
[0125] The second downsampling processing module 402 specifically includes:
[0126] The second downsampling submodule is used to downsample the first fused image according to a preset second downsampling method to obtain the third downsampling image and the fourth downsampling image;
[0127] The second sub-kernel acquisition submodule is used to acquire the third convolutional sub-kernel that matches the third downsampled image, and to acquire the fourth convolutional sub-kernel that matches the fourth downsampled image.
[0128] The second image convolution submodule is used to convolve the third downsampled image using the third convolution subkernel to obtain the third convolution image, and to convolve the fourth downsampled image using the fourth convolution subkernel to obtain the fourth convolution image.
[0129] The second image fusion submodule is used to fuse the third and fourth convolutional images to obtain the second fused image.
[0130] Furthermore, the first upsampling processing module 403 specifically includes:
[0131] The third image convolution submodule is used to obtain the fifth convolution subkernel and the sixth convolution subkernel, and to use the fifth convolution subkernel to convolve the second fused image to obtain the fifth convolution image, and to use the sixth convolution subkernel to convolve the second fused image to obtain the sixth convolution image.
[0132] The first upsampling submodule is used to upsample the fifth convolutional image according to the first upsampling method to obtain the first upsampled image, and to upsample the sixth convolutional image according to the first upsampling method to obtain the second upsampled image;
[0133] The third image fusion submodule is used to fuse the first upsampled image and the second upsampled image to obtain the third fused image.
[0134] The second upsampling processing module 404 specifically includes:
[0135] The fourth image convolution submodule is used to obtain the seventh convolution subkernel and the eighth convolution subkernel, and to use the seventh convolution subkernel to convolve the third fused image to obtain the seventh convolution image, and to use the eighth convolution subkernel to convolve the third fused image to obtain the eighth convolution image.
[0136] The second upsampling submodule is used to upsample the seventh convolutional image according to the second upsampling method to obtain the third upsampled image, and to upsample the eighth convolutional image according to the second upsampling method to obtain the fourth upsampled image;
[0137] The fourth image fusion submodule is used to fuse the third upsampled image and the fourth upsampled image to obtain the fourth fused image.
[0138] Furthermore, the Laplace pyramid image processing device 400 also includes:
[0139] The kernel acquisition module is used to acquire a preset first separable kernel and a second separable kernel;
[0140] The first convolution kernel splitting module is used to split the first separable convolution kernel to obtain a first convolution kernel and a second convolution kernel, wherein both the first convolution kernel and the second convolution kernel are one-dimensional convolution kernels.
[0141] The second convolution kernel splitting module is used to split the second separable convolution kernel to obtain the third convolution kernel and the fourth convolution kernel, wherein the third convolution kernel and the fourth convolution kernel are both one-dimensional convolution kernels.
[0142] Furthermore, the Laplace pyramid image processing device 300 also includes:
[0143] The first convolution kernel downsampling module is used to downsample the first convolution kernel using a first downsampling method to obtain the first convolution sub-kernel and the second convolution sub-kernel.
[0144] The second convolution kernel downsampling module is used to downsample the second convolution kernel using a second downsampling method to obtain the third and fourth convolution sub-kernels.
[0145] The apparatus for processing the Laplace pyramid image also includes:
[0146] The third convolution kernel downsampling module is used to downsample the third convolution kernel using the first downsampling method to obtain the fifth and sixth convolution sub-kernels;
[0147] The fourth convolution kernel downsampling module is used to downsample the fourth convolution kernel using the second downsampling method to obtain the seventh and eighth convolution sub-kernels.
[0148] Furthermore, the first downsampling method is to downsample according to the image pixel columns, and the first downsampling submodule specifically includes:
[0149] The first downsampling unit is used to receive the input image and downsample the odd-numbered columns of the input image to obtain the first downsampled image;
[0150] The second downsampling unit is used to downsample the even-numbered columns of the input image to obtain the second downsampled image;
[0151] The first convolutional kernel downsampling module specifically includes:
[0152] The first convolution kernel downsampling submodule is used to downsample the odd-numbered columns of the first convolution kernel to obtain the first convolution subkernel;
[0153] The second convolution kernel downsampling submodule is used to downsample the even-numbered columns of the first convolution kernel to obtain the second convolution subkernel;
[0154] The third convolution kernel downsampling module specifically includes:
[0155] The fifth convolution kernel downsampling submodule is used to downsample the odd-numbered columns of the third convolution kernel to obtain the fifth convolution kernel;
[0156] The sixth convolution kernel downsampling submodule is used to downsample the even-numbered columns of the third convolution kernel to obtain the sixth convolution kernel;
[0157] The second downsampling method involves downsampling based on image pixel rows. The second downsampling submodule specifically includes:
[0158] The third downsampling unit is used to downsample the odd-numbered rows of the first fused image to obtain the third downsampled image;
[0159] The fourth downsampling unit is used to downsample the even-numbered rows of the first fused image to obtain the fourth downsampled image;
[0160] The second convolutional kernel downsampling module specifically includes:
[0161] The third convolution kernel downsampling submodule is used to downsample the odd-numbered rows of the second convolution kernel to obtain the third convolution kernel;
[0162] The fourth convolution kernel downsampling submodule is used to downsample the even-numbered rows of the second convolution kernel to obtain the fourth convolution subkernel;
[0163] The fourth convolution kernel downsampling module specifically includes:
[0164] The seventh convolution kernel downsampling submodule is used to downsample the odd-numbered rows of the fourth convolution kernel to obtain the seventh convolution kernel;
[0165] The eighth convolution kernel downsampling submodule is used to downsample the even-numbered rows of the fourth convolution kernel to obtain the eighth convolution kernel.
[0166] Furthermore, the first upsampling method is to upsample according to the image pixel columns, and the first upsampling submodule specifically includes:
[0167] The first upsampling unit is used to upsample the even-numbered columns of the fifth convolutional image to obtain the first upsampled image;
[0168] The second upsampling unit is used to upsample the odd columns of the sixth convolutional image to obtain the second upsampled image;
[0169] The second upsampling method involves upsampling based on image pixel rows. The second upsampling submodule specifically includes:
[0170] The third upsampling unit is used to upsample the even-numbered rows of the seventh convolutional image to obtain the third upsampled image;
[0171] The fourth upsampling unit is used to upsample the odd-numbered rows of the eighth convolutional image to obtain the fourth upsampled image.
[0172] In the above embodiments, this application discloses a Laplacian pyramid image processing apparatus, belonging to the field of image processing technology. This application downsamples the input image according to a preset first method, then convolves the downsampled image and fuses the resulting convolutional images; it then downsamples the fused image according to a preset second method, convolves the new downsampled image, and fuses the resulting new convolutional images; it convolves the generated fused image, upsamples the convolutional image according to the preset first method, and fuses the upsampled image; it continues to convolve the generated fused image, upsamples the convolutional image according to the preset second method, and fuses the upsampled image; finally, it subtracts the input image from the final fused image to obtain a difference image. This application significantly reduces computational complexity, reducing chip computing resources and energy consumption while generating the same result.
[0173] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.
Claims
1. A method for processing Laplacian pyramid images, characterized in that, The Laplacian pyramid image processing method is executed by an FPGA chip, and the method includes: Receive an input image, perform a first downsampling on the input image to obtain a first downsampled image and a second downsampled image, perform a first convolution on the first downsampled image to obtain a first convolutional image, and perform a second convolution on the second downsampled image to obtain a second convolutional image, and perform image fusion on the first convolutional image and the second convolutional image to obtain a first fused image; The first fused image is downsampled a second time to obtain a third downsampled image and a fourth downsampled image. The third downsampled image is convolved to obtain a third convolved image, and the fourth downsampled image is convolved to obtain a fourth convolved image. The third convolved image and the fourth convolved image are then fused to obtain a second fused image. The second fused image is subjected to a third convolution to obtain a fifth convolution image and a sixth convolution image. The fifth convolution image is subjected to a first upsampling to obtain a first upsampled image, and the sixth convolution image is subjected to a first upsampling to obtain a second upsampled image. The first upsampled image and the second upsampled image are fused to obtain a third fused image. The third fused image is subjected to a fourth convolution to obtain a seventh convolution image and an eighth convolution image. The seventh convolution image is then upsampled to obtain a third upsampled image, and the eighth convolution image is also upsampled to obtain a fourth upsampled image. The third upsampled image and the fourth upsampled image are then fused to obtain a fourth fused image. The input image and the fourth fused image are subtracted to obtain a difference image, which is then used as the output image of the Laplacian pyramid.
2. The method for processing Laplacian pyramid images as described in claim 1, characterized in that, The process of receiving an input image, performing a first downsampling on the input image to obtain a first downsampling image and a second downsampling image, performing a first convolution on the first downsampling image to obtain a first convolutional image, and performing a second convolution on the second downsampling image to obtain a second convolutional image, and fusing the first convolutional image and the second convolutional image to obtain a first fused image, specifically includes: Receive an input image and downsample the input image according to a preset first downsampling method to obtain a first downsampled image and a second downsampled image; Obtain a first convolutional kernel that matches the first downsampled image, and obtain a second convolutional kernel that matches the second downsampled image; The first downsampled image is convolved using the first convolutional kernel to obtain a first convolutional image, and the second downsampled image is convolved using the second convolutional kernel to obtain a second convolutional image. The first convolutional image and the second convolutional image are fused to obtain a first fused image.
3. The method for processing the Laplacian pyramid image as described in claim 2, characterized in that, The process of performing a second downsampling on the first fused image to obtain a third downsampling image and a fourth downsampling image, convolving the third downsampling image to obtain a third convolutional image, and convolving the fourth downsampling image to obtain a fourth convolutional image, and then fusing the third and fourth convolutional images to obtain a second fused image, specifically includes: The first fused image is downsampled according to a preset second downsampling method to obtain the third downsampled image and the fourth downsampled image; Obtain a third convolutional kernel that matches the third downsampled image, and obtain a fourth convolutional kernel that matches the fourth downsampled image; The third downsampled image is convolved using the third convolutional kernel to obtain the third convolutional image, and the fourth downsampled image is convolved using the fourth convolutional kernel to obtain the fourth convolutional image. The third convolutional image and the fourth convolutional image are fused to obtain a second fused image.
4. The method for processing Laplacian pyramid images as described in claim 3, characterized in that, The process of performing a third convolution on the second fused image to obtain a fifth convolution image and a sixth convolution image, performing a first upsampling on the fifth convolution image to obtain a first upsampled image, and performing a first upsampling on the sixth convolution image to obtain a second upsampled image, and then fusing the first upsampled image and the second upsampled image to obtain a third fused image, specifically includes: Obtain the fifth convolutional sub-kernel and the sixth convolutional sub-kernel, and use the fifth convolutional sub-kernel to convolve the second fused image to obtain the fifth convolutional image, and use the sixth convolutional sub-kernel to convolve the second fused image to obtain the sixth convolutional image; The fifth convolutional image is upsampled according to the first upsampling method to obtain the first upsampled image, and the sixth convolutional image is upsampled according to the first upsampling method to obtain the second upsampled image; The first upsampled image and the second upsampled image are fused to obtain a third fused image.
5. The method for processing Laplacian pyramid images as described in claim 4, characterized in that, The process of performing a fourth convolution on the third fused image to obtain a seventh convolution image and an eighth convolution image, performing a second upsampling on the seventh convolution image to obtain a third upsampled image, and performing a second upsampling on the eighth convolution image to obtain a fourth upsampled image, and then fusing the third upsampled image and the fourth upsampled image to obtain a fourth fused image, specifically includes: Obtain the seventh and eighth convolutional sub-kernels, and use the seventh convolutional sub-kernel to convolve the third fused image to obtain the seventh convolutional image; and use the eighth convolutional sub-kernel to convolve the third fused image to obtain the eighth convolutional image. The seventh convolutional image is upsampled according to the second upsampling method to obtain the third upsampled image, and the eighth convolutional image is upsampled according to the second upsampling method to obtain the fourth upsampled image; The third upsampled image and the fourth upsampled image are fused to obtain a fourth fused image.
6. The method for processing Laplacian pyramid images as described in claim 5, characterized in that, Before receiving the input image and downsampling the input image according to a preset first downsampling method to obtain the first downsampled image and the second downsampled image, the method further includes: Obtain the preset first and second separable convolution kernels; The first separable convolutional kernel is split to obtain a first convolutional kernel and a second convolutional kernel, wherein both the first convolutional kernel and the second convolutional kernel are one-dimensional convolutional kernels; The second separable convolution kernel is split to obtain a third convolution kernel and a fourth convolution kernel, wherein the third convolution kernel and the fourth convolution kernel are both one-dimensional convolution kernels.
7. The method for processing the Laplacian pyramid image as described in claim 6, characterized in that, After splitting the first separable convolutional kernel to obtain the first convolutional kernel and the second convolutional kernel, the method further includes: The first convolutional kernel is downsampled using the first downsampling method to obtain the first convolutional sub-kernel and the second convolutional kernel; The second convolutional kernel is downsampled using the second downsampling method to obtain the third convolutional sub-kernel and the fourth convolutional sub-kernel; After splitting the second separable convolutional kernel to obtain the third and fourth convolutional kernels, the method further includes: The third convolutional kernel is downsampled using the first downsampling method to obtain the fifth and sixth convolutional sub-kernels; The fourth convolutional kernel is downsampled using the second downsampling method to obtain the seventh and eighth convolutional sub-kernels.
8. The method for processing Laplacian pyramid images as described in claim 7, characterized in that, The first downsampling method is to downsample according to the image pixel column. The process of receiving an input image and downsampling the input image according to the preset first downsampling method to obtain the first downsampled image and the second downsampled image specifically includes: Receive the input image, and downsample the odd-numbered columns of the input image to obtain the first downsampled image; The even-numbered columns of the input image are downsampled to obtain the second downsampled image; The step of downsampling the first convolutional kernel using the first downsampling method to obtain the first convolutional sub-kernel and the second convolutional kernel specifically includes: The odd-numbered columns of the first convolutional kernel are downsampled to obtain the first convolutional sub-kernel; The even-numbered columns of the first convolutional kernel are downsampled to obtain the second convolutional sub-kernel; The step of downsampling the third convolutional kernel using the first downsampling method to obtain the fifth and sixth convolutional sub-kernels specifically includes: The odd-numbered columns of the third convolutional kernel are downsampled to obtain the fifth convolutional sub-kernel; The even-numbered columns of the third convolutional kernel are downsampled to obtain the sixth convolutional sub-kernel; The second downsampling method is to downsample according to the image pixel rows. The step of downsampling the first fused image according to the preset second downsampling method to obtain the third downsampled image and the fourth downsampled image specifically includes: The odd-numbered rows of the first fused image are downsampled to obtain the third downsampled image; The even-numbered rows of the first fused image are downsampled to obtain the fourth downsampled image; The step of downsampling the second convolutional kernel using the second downsampling method to obtain the third and fourth convolutional sub-kernels specifically includes: The odd-numbered rows of the second convolutional kernel are downsampled to obtain the third convolutional sub-kernel; The even-numbered rows of the second convolutional kernel are downsampled to obtain the fourth convolutional sub-kernel; The step of downsampling the fourth convolutional kernel using the second downsampling method to obtain the seventh and eighth convolutional sub-kernels specifically includes: The odd-numbered rows of the fourth convolutional kernel are downsampled to obtain the seventh convolutional sub-kernel; The even-numbered rows of the fourth convolutional kernel are downsampled to obtain the eighth convolutional sub-kernel.
9. The method for processing Laplacian pyramid images as described in claim 5, characterized in that, The first upsampling method is to upsample according to the image pixel column. Specifically, upsampling the fifth convolutional image according to the first upsampling method to obtain the first upsampled image, and upsampling the sixth convolutional image according to the first upsampling method to obtain the second upsampled image, includes: Upsample the even-numbered columns of the fifth convolutional image to obtain the first upsampled image; Upsample the odd-numbered columns of the sixth convolutional image to obtain the second upsampled image; The second upsampling method is to upsample according to the image pixel rows. Specifically, upsampling the seventh convolutional image according to the second upsampling method to obtain the third upsampled image, and upsampling the eighth convolutional image according to the second upsampling method to obtain the fourth upsampled image, includes: Upsample the even-numbered rows of the seventh convolutional image to obtain the third upsampled image; The odd-numbered rows of the eighth convolutional image are upsampled to obtain the fourth upsampled image.
10. An apparatus for processing Laplacian pyramid images, characterized in that, The Laplacian pyramid image processing device is housed in an FPGA chip, and the device includes: The first downsampling processing module is used to receive an input image, perform a first downsampling on the input image to obtain a first downsampling image and a second downsampling image, perform a first convolution on the first downsampling image to obtain a first convolutional image, and perform a second convolution on the second downsampling image to obtain a second convolutional image, and perform image fusion on the first convolutional image and the second convolutional image to obtain a first fused image. The second downsampling processing module is used to perform a second downsampling on the first fused image to obtain a third downsampling image and a fourth downsampling image, to perform convolution on the third downsampling image to obtain a third convolutional image, and to perform convolution on the fourth downsampling image to obtain a fourth convolutional image, and to perform image fusion on the third convolutional image and the fourth convolutional image to obtain a second fused image. The first upsampling processing module is used to perform a third convolution on the second fused image to obtain a fifth convolution image and a sixth convolution image, perform a first upsampling on the fifth convolution image to obtain a first upsampled image, and perform a first upsampling on the sixth convolution image to obtain a second upsampled image, and perform image fusion on the first upsampled image and the second upsampled image to obtain a third fused image. The second upsampling processing module is used to perform a fourth convolution on the third fused image to obtain a seventh convolution image and an eighth convolution image, perform a second upsampling on the seventh convolution image to obtain a third upsampled image, and perform a second upsampling on the eighth convolution image to obtain a fourth upsampled image, and perform image fusion on the third upsampled image and the fourth upsampled image to obtain a fourth fused image. The image subtraction module is used to perform a subtraction operation on the input image and the fourth fused image to obtain a difference image, and use the difference image as the output image of the Laplacian pyramid.