WO2023015871A1 - 图像特征提取方法及装置、计算机设备和可读存储介质 - Google Patents

图像特征提取方法及装置、计算机设备和可读存储介质 Download PDF

Info

Publication number
WO2023015871A1
WO2023015871A1 PCT/CN2022/078874 CN2022078874W WO2023015871A1 WO 2023015871 A1 WO2023015871 A1 WO 2023015871A1 CN 2022078874 W CN2022078874 W CN 2022078874W WO 2023015871 A1 WO2023015871 A1 WO 2023015871A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
target
convolution kernel
map
feature map
Prior art date
Application number
PCT/CN2022/078874
Other languages
English (en)
French (fr)
Inventor
叶涛
王圻
朱江瀚
王世航
王星博
Original Assignee
南方科技大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 南方科技大学 filed Critical 南方科技大学
Publication of WO2023015871A1 publication Critical patent/WO2023015871A1/zh
Priority to US18/396,324 priority Critical patent/US20240185571A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/22—Image preprocessing by selection of a specific region containing or referencing a pattern; Locating or processing of specific regions to guide the detection or recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/7715—Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Definitions

  • the present application relates to the technical field of convolutional neural networks, and in particular, relates to an image feature extraction method and device, computer equipment, and a readable storage medium.
  • the purpose of this application is to provide an image feature extraction method and device, computer equipment and readable storage medium, which can reduce the consumption of computing resources during feature extraction and improve feature extraction efficiency.
  • the present application provides an image feature extraction method, the method comprising:
  • the convolution operation is performed on the feature subgraph to be convoluted using the target convolution kernel based on the Karatsuba algorithm to obtain a corresponding target feature subgraph;
  • the target feature submaps corresponding to adjacent positions are boundary superimposed to obtain an overlapping feature map ;
  • boundary circumcision is performed on the overlapping feature map to obtain a corresponding output feature map, wherein the output feature map is used to represent the The feature extraction result of the input feature map.
  • the convolution kernel size of the target convolution kernel calculate the first target number of blank element rings required for a single input feature subgraph, wherein the first target number is determined by the row/column corresponding to the convolution kernel size subtract one from the number;
  • a first target number of blank element rings are sequentially filled around the input feature subgraph with the input feature subgraph as the center to obtain a corresponding feature subgraph to be convoluted.
  • the target convolution kernel is used to perform convolution operations on the feature subgraphs to be convolved based on the Karatsuba algorithm to obtain the corresponding target feature subgraphs Figure steps, including:
  • the row and column feature expressions of the input feature subgraph are constructed based on the Karatsuba algorithm
  • the rank and column feature expressions of the target convolution kernel are constructed based on the Karatsuba algorithm
  • the multiplication process of the row and column feature expressions of the corresponding input feature subgraph and the row and column feature expressions of the target convolution kernel is split and converted based on the Karatsuba algorithm to obtain the corresponding Split conversion results;
  • the corresponding split conversion results are expressed elementally based on the preset convolution map distribution corresponding to the target convolution kernel, to obtain the corresponding target feature subgraphs.
  • the boundary superposition of the target feature submaps corresponding to adjacent positions is performed to obtain the overlapping feature map.
  • the convolution kernel size of the target convolution kernel determine the second target number of the to-be-superimposed element ring of a single target feature subgraph, wherein the second target number is determined by the row/column corresponding to the convolution kernel size subtract one from the number;
  • For each target feature sub-graph select the second target number of sub-graph element rings of the target feature sub-graph from outside to inside as element rings to be superimposed, and obtain the area to be superimposed of the target feature sub-graph;
  • image stitching is performed on each target feature submap to obtain a corresponding image stitching result
  • the overlapping feature map is obtained by elementally superimposing the respective contact edge regions of adjacent to-be-superimposed regions.
  • the step of performing boundary circumcision on the overlapping feature map according to the convolution kernel size of the target convolution kernel to obtain a corresponding output feature map includes:
  • the convolution kernel size of the target convolution kernel determine the third target number of the element ring to be cut off in the overlapping feature map, wherein the third target number is determined by the row/column corresponding to the convolution kernel size subtract one from the number;
  • Image cutting is performed on the region to be cut in the overlapping feature map to obtain the output feature map corresponding to the input feature map.
  • the method also includes:
  • the preset image size is configured according to the convolution kernel size of the target convolution kernel, wherein the number of rows corresponding to the preset image size is greater than or equal to the number of rows corresponding to the convolution kernel size, so The number of columns corresponding to the preset image size is greater than or equal to the number of columns corresponding to the convolution kernel size.
  • the present application provides an image feature extraction device, the device comprising:
  • the feature map acquisition module is used to obtain the feature map to be processed for the target convolution kernel of the feature extraction convolutional neural network
  • a feature map preprocessing module configured to perform image preprocessing on the feature map to be processed according to a preset image size to obtain a corresponding input feature map, wherein the image size of the input feature map is an integer of the preset image size times;
  • An input feature cutting module configured to cut the input feature map into a plurality of input feature submaps according to a preset image size, wherein the image size of each input feature submap is consistent with the preset image size;
  • a subgraph boundary filling module configured to perform boundary filling on the input feature subgraph according to the convolution kernel size of the target convolution kernel for each of the input feature subgraphs, to obtain a corresponding feature subgraph to be convolved;
  • a feature convolution operation module for each of the feature subgraphs to be convoluted, using the target convolution kernel based on the Karatsuba algorithm to perform convolution operations on the feature subgraphs to be convolved, to obtain the corresponding target feature subgraphs;
  • the subgraph boundary superposition module is configured to, for the obtained plurality of target feature subgraphs, combine the target feature subgraphs correspondingly adjacent to each other according to the distribution of each input feature subgraph in the input feature map Perform boundary superposition to obtain overlapping feature maps;
  • the output feature extraction module is used to perform boundary circumcision on the overlapping feature map according to the convolution kernel size of the target convolution kernel to obtain a corresponding output feature map, wherein the output feature map is used to represent the
  • the feature extraction result of the input feature map is processed for feature map matching.
  • the device also includes:
  • a cutting size configuration module configured to configure the preset image size according to the convolution kernel size of the target convolution kernel, wherein the number of rows corresponding to the preset image size is greater than or equal to the convolution kernel size The number of rows corresponding to the preset image size is greater than or equal to the number of columns corresponding to the size of the convolution kernel.
  • the present application provides a computer device, including a processor and a memory, the memory stores a computer program that can be executed by the processor, and the processor can execute the computer program to implement the above-mentioned embodiment.
  • a computer device including a processor and a memory
  • the memory stores a computer program that can be executed by the processor
  • the processor can execute the computer program to implement the above-mentioned embodiment.
  • the present application provides a readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the image feature extraction method described in any one of the foregoing embodiments is implemented.
  • the application After obtaining the to-be-processed feature map of the target convolution kernel of the feature extraction convolutional neural network, the application will cut the input feature map obtained by image preprocessing from the to-be-processed feature map into multiple input features according to the preset image size Subgraph, and fill the boundary of each input feature subgraph to obtain the corresponding feature subgraph to be convoluted, and then for each feature subgraph to be convolved, use the Karatsuba algorithm to reduce the target convolution kernel and the feature to be convolved Multiplier resource loss in the convolution operation process of the sub-graph, and effectively suppress the increase in the resource loss of the adder in the convolution operation process, and obtain the corresponding target feature sub-graph, and then according to each input feature sub-graph in the input feature map Distribution status, the boundary superposition of the target feature submaps corresponding to the adjacent positions is performed to obtain the overlapping feature map, and finally the boundary circumcision is performed on the overlapping feature map to obtain the corresponding output feature map, thereby reducing the calculation of
  • FIG. 1 is a schematic diagram of the composition of computer equipment provided by the embodiment of the present application.
  • Fig. 2 is one of the flow diagrams of the image feature extraction method provided by the embodiment of the present application.
  • FIG. 3 is a schematic flow chart of substeps included in step S240 in FIG. 2;
  • FIG. 4 is a schematic diagram of the distribution of feature subgraphs to be convolved provided by the embodiment of the present application.
  • FIG. 5 is a schematic flow chart of substeps included in step S250 in FIG. 2;
  • FIG. 6 is one of the schematic diagrams showing the distribution status of the preset convolution map provided by the embodiment of the present application.
  • FIG. 7 is the second schematic diagram showing the distribution status of the preset convolution mapping in the embodiment of the present application.
  • FIG. 8 is a schematic flow chart of substeps included in step S260 in FIG. 2;
  • FIG. 9 is a schematic diagram showing the image stitching results provided by the embodiment of the present application.
  • FIG. 10 is a schematic flow chart of substeps included in step S270 in FIG. 2;
  • FIG. 11 is a schematic diagram of the distribution of overlapping feature maps provided by the embodiment of the present application.
  • FIG. 12 is the second schematic flow diagram of the image feature extraction method provided by the embodiment of the present application.
  • FIG. 13 is one of the schematic diagrams of the composition of the image feature extraction device provided by the embodiment of the present application.
  • FIG. 14 is the second schematic diagram of the composition of the image feature extraction device provided by the embodiment of the present application.
  • Icons 10-computer equipment; 11-memory; 12-processor; 13-communication unit; 100-image feature extraction device; 110-feature map acquisition module; 120-feature map preprocessing module; 130-input feature cutting module; 140-subgraph boundary filling module; 150-feature convolution operation module; 160-subgraph boundary superposition module; 170-output feature extraction module; 180-cutting size configuration module.
  • the fast Fourier transform (FFT) or Winograd algorithm can be used to optimize the convolution operation, so as to reduce the original required computing resource consumption of the convolutional neural network , so as to improve the operation efficiency.
  • FFT fast Fourier transform
  • Winograd algorithm can be used to optimize the convolution operation, so as to reduce the original required computing resource consumption of the convolutional neural network , so as to improve the operation efficiency.
  • the convolution optimization process implemented by the FFT algorithm is to transform the convolution operation in the time domain into the product operation in the comment, which involves complex number operations, and one complex number multiplication is equivalent to four real number multiplication and addition operations , which greatly increases the memory bandwidth requirements of the convolutional neural network when it is running. On the whole, the multiplier resource loss and the adder resource loss are still high, and the convolution optimization effect is not good.
  • the convolution optimization process implemented by the Winograd algorithm is to map the data to another real number space, and use addition and subtraction instead of part of the multiplication operation, so that "the addition and subtraction operation speed is much higher than the multiplication operation, and the multiplier is running.
  • the calculation resource consumption required is much higher than the calculation resource consumption required by the adder when it is running to improve the operation efficiency. It often requires a large amount of adder resource consumption to complete the replacement of part of the multiplication operation, so this method essentially has addition. However, the overall convolution optimization effect still appears to be poor.
  • the embodiment of the present application is realized by providing an image feature extraction method and device, computer equipment, and a readable storage medium.
  • FIG. 1 is a schematic composition diagram of a computer device 10 provided by an embodiment of the present application.
  • the computer device 10 is used to quickly realize the image feature extraction operation, wherein the computer device 10 stores a feature extraction convolutional neural network, and can extract the image to be processed based on the feature extraction convolutional neural network
  • the convolution operation process of the feature extraction convolutional neural network is effectively optimized, so that the resource loss of the multiplier corresponding to the original convolution operation process is effectively reduced, and the original convolution operation is effectively suppressed synchronously
  • the increase of the resource loss of the adder corresponding to the process reduces the computational resource loss during feature extraction as a whole, improves the efficiency of feature extraction, and improves the convolution optimization effect of the neural network.
  • the computer device 10 may be, but not limited to, a smart phone, a tablet computer, a personal computer, a server, and the like.
  • the computer device 10 may include a memory 11 , a processor 12 , a communication unit 13 and an image feature extraction device 100 .
  • the components of the memory 11 , the processor 12 and the communication unit 13 are directly or indirectly electrically connected to each other to realize data transmission or interaction.
  • the memory 11 , the processor 12 and the communication unit 13 can be electrically connected to each other through one or more communication buses or signal lines.
  • the memory 11 can be, but not limited to, Random Access Memory (Random Access Memory, RAM), Read Only Memory (Read Only Memory, ROM), Programmable Read-Only Memory (Programmable Read-Only Memory, PROM), Erasable Programmable Read-Only Memory (EPROM), Electric Erasable Programmable Read-Only Memory (EEPROM), etc.
  • RAM Random Access Memory
  • ROM Read Only Memory
  • PROM Programmable Read-Only Memory
  • PROM Programmable Read-Only Memory
  • EPROM Erasable Programmable Read-Only Memory
  • EEPROM Electric Erasable Programmable Read-Only Memory
  • the memory 11 is also used to store the feature extraction convolutional neural network, so as to use the feature extraction convolutional neural network to realize the image feature extraction operation, wherein the feature extraction convolutional neural network may include at least one convolutional layer, each The convolutional layer may include at least one special convolution kernel for extracting image features, and the special convolution kernel may be a 3*3 convolution kernel or a 5*5 convolution kernel.
  • the processor 12 may be an integrated circuit chip capable of processing signals.
  • Described processor 12 can be general-purpose processor, comprises central processing unit (Central Processing Unit, CPU), graphics processing unit (Graphics Processing Unit, GPU) and network processor (Network Processor, NP), digital signal processor (DSP) ), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.
  • the general-purpose processor may be a microprocessor, or the processor may be any conventional processor, etc., and may realize or execute various methods, steps, and logic block diagrams disclosed in the embodiments of the present application.
  • the communication unit 13 is used to establish a communication connection between the computer device 10 and other electronic devices through a network, and send and receive data through the network, wherein the network includes a wired communication network and a wireless communication network.
  • the computer device 10 may be communicatively connected to a smart phone through the communication unit 13, so as to acquire the image to be processed requiring feature extraction from the smart phone.
  • the image feature extraction apparatus 100 includes at least one software function module that can be stored in the memory 11 or in the operating system of the computer device 10 in the form of software or firmware.
  • the processor 12 can be used to execute executable modules stored in the memory 11 , such as software function modules and computer programs included in the image feature extraction device 100 .
  • the computer device 10 can effectively optimize the convolution operation process of the feature extraction convolution neural network during the feature extraction process of the image to be processed using the feature extraction convolution neural network through the image feature extraction device 100, so that The resource loss of the multiplier corresponding to the original convolution operation process is effectively reduced, and the increase of the resource loss of the adder corresponding to the original convolution operation process is effectively suppressed simultaneously, thereby reducing the computing resource loss during feature extraction as a whole, and Improve the efficiency of feature extraction and improve the convolution optimization effect of the neural network.
  • FIG. 1 is only a schematic composition diagram of the computer device 10, and the computer device 10 may also include more or less components than those shown in FIG. 1 shows different configurations. Each component shown in Fig. 1 may be implemented by hardware, software or a combination thereof.
  • the convolution optimization of the neural network is improved.
  • the embodiment of the present application realizes the foregoing functions by providing an image feature extraction method, and the image feature extraction method provided in the present application will be described in detail below.
  • FIG. 2 is one of the schematic flowcharts of the image feature extraction method provided by the embodiment of the present application.
  • the image feature extraction method shown in FIG. 2 may include steps S210 to S260.
  • Step S210 acquiring the feature map to be processed of the target convolution kernel of the feature extraction convolutional neural network.
  • the target convolution kernel is a convolution kernel in the feature extraction convolutional neural network that needs to implement feature extraction operations
  • the convolution kernel size of the target convolution kernel can be, but not limited to , 3*3, 5*5, etc.
  • the feature map to be processed is a feature image that needs to use the target convolution kernel to implement a feature extraction operation.
  • step S220 image preprocessing is performed on the feature map to be processed according to the preset image size to obtain a corresponding input feature map.
  • the preset image size is used to represent the standard size of the input feature map when it is cut, so when the computer device 10 obtains the feature map to be processed, it will be based on the preset image size.
  • Image preprocessing is performed on the feature map to be processed, so that the image size of the input feature map obtained by image preprocessing is an integer multiple of the preset image size.
  • the image preprocessing will correspondingly include an image filling operation, which can be used in the image size of the feature map to be processed Blank elements with a padding value of zero around the boundary or at a side edge position make the final image size of the input feature map an integer multiple of the preset image size.
  • the image preprocessing may also include any one of image digitization processing, image geometric transformation processing, image normalization processing, image smoothing processing, image restoration processing, and image enhancement. one or more combinations.
  • Step S230 cutting the input feature map into multiple input feature sub-maps according to the preset image size.
  • the preset image size is used to represent a standard size when the input feature map is cut, and the sub-image size of each input feature sub-map is consistent with the preset image size.
  • the preset image size usually needs to be equal to or larger than the convolution kernel size of the target convolution kernel, that is, the number of image element rows corresponding to the preset image size is greater than or equal to the volume of the target convolution kernel.
  • the number of convolution element rows corresponding to the kernel size, the number of image element columns corresponding to the preset image size is greater than or equal to the convolution element column number corresponding to the convolution kernel size of the target convolution kernel.
  • the input feature map can be obtained by splicing the cut out multiple input feature submaps.
  • Step S240 for each input feature subgraph, perform boundary padding on the input feature subgraph according to the convolution kernel size of the target convolution kernel, to obtain the corresponding feature subgraph to be convolved.
  • the feature subgraph to be convolved is used to represent a feature subgraph that can directly perform a convolution operation with the target convolution kernel.
  • the computer device 10 After the computer device 10 obtains a plurality of input feature subgraphs corresponding to the input feature map, it can perform boundary padding on a single input feature subgraph according to the convolution kernel size of the corresponding target convolution kernel, so as to ensure that the obtained to-be-convoluted
  • the convolution feature subgraph and the target convolution kernel are adapted to each other within the convolution operation dimension.
  • FIG. 3 is a schematic flowchart of the substeps included in step S240 in FIG. 2
  • FIG. 4 is a schematic diagram of the distribution of feature subgraphs to be convolved provided by the embodiment of the present application.
  • the step S240 may include substeps S241 to S242.
  • Sub-step S241 according to the convolution kernel size of the target convolution kernel, calculate the first target number of blank element rings required for a single input feature subgraph, wherein the first target number is subtracted from the number of rows/columns corresponding to the convolution kernel size one get.
  • the blank element ring is used to represent that a plurality of blank elements included in the blank element ring are close to each other and distributed in a hollow ring shape, and the element value of the blank element is zero.
  • the first target number is used to indicate the number of blank element rings that need to be filled when the corresponding input feature subgraph realizes boundary filling, and the first target number is corresponding to the convolution kernel size of the target convolution kernel Get by subtracting one from the number of rows/columns.
  • Sub-step S242 for each input feature subgraph, fill the first target number of blank element rings sequentially with the input feature subgraph as the center and expand around the input feature subgraph to obtain the corresponding feature subgraph to be convoluted .
  • the computer device 10 determines the number of blank element rings (that is, the first target number) that need to be filled in a single input feature subgraph, it will correspondingly set boundaries around the corresponding input feature subgraph Filling the first target number of blank element rings formulaically to form a corresponding feature subgraph to be convolved, wherein if multiple blank element rings are filled, the number of elements in the adjacent two blank element rings is smaller
  • the blank element ring is embedded in the hollow part of the blank element ring with more elements.
  • each The solid grid represents the subgraph element of an input feature subgraph) whose subgraph size is also 3*3
  • the input feature subgraph Centering around the boundary of the input feature subgraph fill in two blank element rings sequentially, so that the two blank element rings are nested together, the closer to the input feature subgraph, the greater the number of elements in the blank element ring.
  • the present application can ensure that the constructed feature subgraph to be convolved and the corresponding target convolution kernel are compatible within the convolution operation dimension by executing the above sub-step S241 and sub-step S242.
  • Step S250 for each feature subgraph to be convolved, perform a convolution operation on the feature subgraph to be convolved using the target convolution kernel based on the Karatsuba algorithm to obtain a corresponding target feature subgraph.
  • the computer device 10 can perform convolution operation on the feature subgraph to be convolved with the target convolution kernel
  • the Karatsuba algorithm is used for correlation processing to obtain the input feature subgraph corresponding to the input feature subgraph that completes the convolution operation, so that a small amount of adder resource loss is used to replace part of the original multiplication in the convolution operation process through the Karatsuba algorithm calculation, to reduce the multiplier resource loss during the convolution operation of the target convolution kernel and the feature subgraph to be convoluted, and effectively suppress the increase of the adder resource loss during the convolution operation.
  • the Karatsuba algorithm can split the multiplication operation of the participating multiplier and the multiplicand into three partial product operations based on the divide-and-conquer idea, thereby reducing the multiplication computing resource consumption required by the original product operation, and can use a small amount of addition Computational resource consumption achieves corresponding multiplication calculation effects.
  • FIG. 5 is a schematic flowchart of the sub-steps included in step S250 in FIG. 2, and FIG. Showing one of the schematic diagrams
  • FIG. 7 is the second schematic diagram showing the distribution status of the preset convolution map in the embodiment of the present application.
  • the step S250 may include substeps S251 to S254.
  • Sub-step S251 for each feature subgraph to be convoluted, according to the characteristic element values in different distribution positions in the corresponding input feature subgraph, the row and column feature expressions of the input feature subgraph are constructed based on the Karatsuba algorithm.
  • the row and column feature expressions of the feature subgraph to be convolved are used to represent effective image features corresponding to the feature subgraph to be convolved that match the input feature subgraph.
  • the effective image features corresponding to the input feature subgraph can be regarded as the value formed by the feature element values in different distribution positions according to the row and column distribution.
  • Sub-step S252 for the target convolution kernel, according to the convolution element values at different distribution positions in the target convolution kernel, construct the rank and column feature expressions of the target convolution kernel based on the Karatsuba algorithm.
  • the row and column feature expressions of the target convolution kernel are used to represent the effective image features of the corresponding target convolution kernel.
  • the effective image features corresponding to the target convolution kernel can be regarded as the value formed by the convolution element values at different distribution positions according to the row and column distribution.
  • the row feature expression of the element row is constructed using the Karatsuba algorithm, and the row feature expression uses each convolution element value of the element row as the element row expression weight , and take the distribution digits of each convolution element value as the row feature expression power, and use the row feature expression base to perform multiplication and addition to form. Then, use the row feature expressions of different element rows as the element column expression weights of the target convolution kernel, and use the distribution digits of the row feature expressions of each element row as the column feature expression power, and use the column feature expression base to perform Multiplication and addition are combined to obtain the row-column feature expression of the target convolution kernel.
  • the effective features corresponding to the two-dimensional reference convolution kernel are For non-blank elements b1, b2, b3, b4, b5, b6, b7, b8, and b9, the first three non-blank elements are in the same element row, the middle three non-blank elements are in the same element row, and the last three non-blank elements are in the same element row
  • Sub-step S253 for each feature subgraph to be convolved, split and convert the multiplication process of the row and column feature expressions of the corresponding input feature subgraph and the row and column feature expressions of the target convolution kernel based on the Karatsuba algorithm to obtain the corresponding Split conversion results.
  • the row and column feature expressions of the feature subgraph to be convoluted can be combined with the row and column feature expressions of the target convolution kernel Multiplication between expressions
  • the weight coefficients in the expression are split and converted to achieve the corresponding multiplication calculation effect by using a small amount of addition calculation resource consumption, reducing the multiplication calculation resource consumption required by the original product operation, and obtaining the corresponding Split conversion results.
  • the present application can realize the corresponding multiplication calculation effect with only a small amount of addition calculation resource consumption by executing the above sub-step S253, reduce the multiplication calculation resource consumption required by the original product operation, and obtain the corresponding split conversion result.
  • Sub-step S254 for each feature subgraph to be convoluted, express the corresponding split conversion result based on the preset convolution map distribution corresponding to the target convolution kernel, and obtain the corresponding target feature subgraph.
  • the preset convolution map distribution status is used to indicate that in the feature map results obtained through correlation convolution operations based on the mutually adapted reference feature map and the reference convolution kernel, the difference in the feature map results is
  • the calculated mapping relationship between the characteristic element value of the position and each characteristic element value in the reference feature map and each convolution element value in the reference convolution kernel for example, the distribution of the one-dimensional preset convolution map in Figure 6 and the one-dimensional reference
  • the feature map and the one-dimensional reference convolution kernel are adapted to each other, and the distribution of the two-dimensional preset convolution map in Figure 7 is adapted to the two-dimensional reference feature map and the two-dimensional reference convolution kernel.
  • the computer device 10 After the computer device 10 obtains a split conversion result corresponding to a feature subgraph to be convoluted and matched with the target convolution kernel, it can use a pre-convolution image whose format matches the feature subgraph to be convoluted and the target convolution kernel. Set the distribution of the convolution map to construct the feature map of the split conversion result, and express the specific element values of each element in the constructed feature map to obtain the target feature submap corresponding to the feature submap to be convolved.
  • the present application can use a small amount of adder resource loss to replace part of the original multiplication calculation in the convolution operation process by executing the above substep S251 to substep S254, so as to reduce the target convolution kernel and the feature to be convoluted Multiplier resource loss during the convolution operation of the graph, and effectively suppress the increase in the resource loss of the adder during the convolution operation.
  • Step S260 for the obtained multiple target feature submaps, according to the distribution of each input feature submap in the input feature map, perform boundary superposition of correspondingly adjacent target feature submaps to obtain an overlapping feature map.
  • these multiple target feature subgraphs can be boundary superimposed to achieve Feature concatenation operation on multiple target feature subgraphs.
  • FIG. 8 is a schematic flowchart of the sub-steps included in step S260 in FIG. 2
  • FIG. 9 is a schematic diagram showing the image stitching result provided by the embodiment of the present application.
  • the step S260 may include substeps S261 to S264.
  • Sub-step S261 according to the convolution kernel size of the target convolution kernel, determine the second target number of the to-be-superimposed element ring of a single target feature subgraph, wherein the second target number is reduced by the number of rows/columns corresponding to the convolution kernel size one get.
  • the element ring to be superimposed is used to represent a plurality of feature elements that are distributed in a hollow ring in the corresponding target feature sub-map and need to be superimposed with other target feature sub-maps.
  • the second target number is used to indicate the number of element rings to be superimposed near the boundary of the corresponding target feature subgraph, wherein among the adjacent element rings to be superimposed in the same target feature subgraph, the element rings to be superimposed that are far away from the boundary are replaced by the element rings that are close to the boundary Elements to be superimposed are nested in a ring, and the second target number is obtained by subtracting one from the number of rows/columns corresponding to the size of the convolution kernel of the target convolution kernel.
  • Sub-step S262 for each target feature sub-graph, select the second target number of sub-graph element rings of the target feature sub-graph from outside to inside as element rings to be superimposed, and obtain the area to be superimposed of the target feature sub-graph.
  • the computer device 10 determines the number of element rings to be superimposed (i.e., the second target number) required to be superimposed on a single target feature subgraph, it will correspondingly start from the boundary of the target feature subgraph from the outer Inwardly select a second target number of subgraph element rings of the target feature subgraph as element rings to be superimposed, and combine all selected element rings to be superimposed to form a region to be superimposed of the target feature subgraph.
  • the sub-graph element ring is used to represent a plurality of closely adjacent sub-graph elements distributed in a hollow ring in the corresponding target feature sub-graph.
  • each solid grid represents a subgraph element in the non-overlapping area of a target feature submap.
  • image stitching is performed on each target feature sub-map to obtain a corresponding image stitching result.
  • target feature submaps when multiple target feature submaps are spliced together, they will be arranged according to the distribution position of the input feature submap corresponding to the target feature submap in the input feature map, so that the corresponding obtained image
  • the splicing result is similar to the input feature map in terms of feature position distribution, as shown in Figure 9, the image splicing result obtained by splicing the four target feature submaps.
  • Sub-step S264 in the image mosaic result, elementally superimpose the respective contact edge regions of adjacent to-be-superimposed regions to obtain a superimposed feature map.
  • the contact edge region on each region to be superimposed is the The area range between the edge position of the region in contact with other regions to be superimposed and the non-superimposed region in the target feature submap, such as in the two adjacent target feature submaps at the top of Figure 9, the upper left target feature submap
  • the contact edge area of the map corresponds to the area enclosed by the vertical dotted line box in the area to be superimposed
  • the contact edge area of the target feature submap on the upper right corresponds to the area enclosed by the vertical dotted line box in the area to be superimposed range
  • the respective contact edge regions of the two target feature submaps are in contact with each other.
  • element superposition is performed on the respective contact edge regions of the adjacent to-be-superimposed regions in the image mosaic result, so as to realize the feature mosaic operation of the respective sub-image features of multiple target feature sub-graphs in the image mosaic result, and obtain the results including An overlapping feature map that can characterize the feature extraction result of the input feature map with interference noise.
  • the present application can implement the feature splicing operation on multiple target feature submaps by executing the above substeps S261 to substep S264, and obtain the overlapping of feature extraction results that can characterize the input feature map including interference noise feature map.
  • step S270 boundary circumcision is performed on the overlapped feature map according to the convolution kernel size of the target convolution kernel to obtain a corresponding output feature map.
  • the output feature map is used to represent the feature extraction result of the input feature map matched with the feature map to be processed, and the boundary ring cutting operation can remove the interference noise in the overlapping feature map , retaining the final feature extraction result matching the input feature map.
  • FIG. 10 is a schematic flowchart of the sub-steps included in step S270 in FIG. 2
  • FIG. 11 is a schematic diagram of the distribution of overlapping feature maps provided by the embodiment of the present application.
  • the step S270 may include substeps S271 to S273.
  • Sub-step S271 according to the convolution kernel size of the target convolution kernel, determine the third target number of the element ring to be removed in the overlapping feature map, where the third target number is obtained by subtracting one from the number of rows/columns corresponding to the convolution kernel size .
  • the element ring to be circumscribed is used to represent a plurality of feature elements that need to be removed in a hollow circular distribution in the overlapping feature map
  • the third target number is used to represent the proximity of the corresponding overlapping feature map.
  • Sub-step S272 select a third target number of feature map element rings of the overlapping feature map from outside to inside as element rings to be removed, and obtain the region to be removed of the overlapping feature map.
  • the computer device 10 after determining the number of third objects corresponding to the overlapping feature map, the computer device 10 will correspondingly select the third object of the overlapping feature map from outside to inside from the boundary of the overlapping feature map
  • a number of feature map element rings are used as element rings to be cut off, and all selected element rings to be cut off are combined to form a cut-off area of the overlapping feature map.
  • the feature map element ring is used to represent a plurality of feature map elements close to each other that are distributed in a hollow ring in the corresponding overlapping feature map.
  • sub-step S273 image cutting is performed on the region to be cut in the overlapping feature map to obtain an output feature map corresponding to the input feature map.
  • the present application can remove the interference noise in the overlapping feature map by performing the above sub-step S271 to sub-step S273, and retain the final feature extraction result that matches the input feature map.
  • the present application is able to reduce the multiplier resource loss required for the convolution operation process in the image feature extraction process, and effectively suppress the adder resource loss required for the convolution operation process by performing the above steps S210 to S270 In order to reduce the consumption of computing resources during feature extraction as a whole, improve the efficiency of feature extraction, and improve the convolution optimization effect of feature extraction convolutional neural network.
  • FIG. 12 is a second schematic flowchart of an image feature extraction method provided by an embodiment of the present application.
  • the image feature extraction method shown in FIG. 12 may further include step S209 .
  • Step S209 configuring the preset image size according to the convolution kernel size of the target convolution kernel.
  • the configured number of rows corresponding to the preset image size is greater than or equal to the number of convolution element rows corresponding to the convolution kernel size of the target convolution kernel, and the configured The number of image element columns corresponding to the preset image size is greater than or equal to the number of convolution element columns corresponding to the convolution kernel size of the target convolution kernel.
  • the present application can ensure that the cut out multiple input feature subgraphs are spliced together by performing the step S209 to obtain an input feature map that matches the feature map to be processed, and the Karatsuba algorithm can be used to make the corresponding input feature subgraph Perform convolution operation with the target convolution kernel.
  • this application implements the aforementioned functions by dividing the image feature extraction device 100 into functional modules.
  • the specific composition of the image feature extraction device 100 provided in this application will be described accordingly below.
  • the image feature extraction device 100 may include a feature map acquisition module 110, a feature map preprocessing module 120, an input feature cutting module 130, a sub-graph boundary filling module 140, a feature convolution operation module 150, a sub-image Graph boundary overlay module 160 and output feature extraction module 170 .
  • the feature map acquisition module 110 is configured to acquire the feature map to be processed for the target convolution kernel of the feature extraction convolutional neural network.
  • the feature map preprocessing module 120 is configured to perform image preprocessing on the feature map to be processed according to a preset image size to obtain a corresponding input feature map, wherein the image size of the input feature map is the preset image size Integer multiple;
  • the input feature cutting module 130 is configured to cut the input feature map into a plurality of input feature submaps according to the preset image size, wherein the image size of each input feature submap remains the same as the preset image size unanimous.
  • the subgraph boundary filling module 140 is configured to, for each of the input feature subgraphs, perform boundary filling on the input feature subgraph according to the convolution kernel size of the target convolution kernel, to obtain the corresponding feature subgraph to be convolved .
  • the feature convolution operation module 150 is used to perform a convolution operation on each of the feature subgraphs to be convoluted using the target convolution kernel based on the Karatsuba algorithm to obtain a corresponding target feature subgraph .
  • the subgraph boundary superposition module 160 is configured to, for the obtained plurality of target feature subgraphs, according to the distribution status of each of the input feature subgraphs in the input feature map, combine the target feature subgraphs corresponding to adjacent positions The graphs are superimposed on the boundaries to obtain overlapping feature maps.
  • the output feature extraction module 170 is configured to perform boundary circumcision on the overlapping feature map according to the convolution kernel size of the target convolution kernel to obtain a corresponding output feature map, wherein the output feature map is used to represent the The feature extraction result of the input feature map to be processed for feature map matching.
  • FIG. 14 is a second schematic diagram of the composition of the image feature extraction device 100 provided by the embodiment of the present application.
  • the image feature extraction apparatus 100 may further include a cutting size configuration module 180 .
  • the cutting size configuration module 180 is configured to configure the preset image size according to the convolution kernel size of the target convolution kernel, wherein the number of rows corresponding to the preset image size is greater than or equal to the convolution kernel The number of rows corresponding to the size, the number of columns corresponding to the preset image size is greater than or equal to the number of columns corresponding to the size of the convolution kernel.
  • the above-mentioned elements may be the image pixels of the image, or the image in a certain color channel (for example, R (Red, red) channel, G (Green, green) channel or B (Blue, blue) channel), the specific element expression can be adaptively adjusted according to the image feature extraction requirements.
  • R Red, red
  • G Green, green
  • B Blue, blue
  • each block in a flowchart or block diagram may represent a module, program segment, or part of code that includes one or more Executable instructions.
  • the functions noted in the block may occur out of the order noted in the figures. For example, two blocks in succession may, in fact, be executed substantially concurrently, or they may sometimes be executed in the reverse order, depending upon the functionality involved.
  • each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations can be implemented by a dedicated hardware-based system that performs the specified function or action , or may be implemented by a combination of dedicated hardware and computer instructions.
  • each functional module in each embodiment of the present application may be integrated to form an independent part, each module may exist independently, or two or more modules may be integrated to form an independent part. If the functions are realized in the form of software function modules and sold or used as independent products, they can be stored in a readable storage medium.
  • the technical solution of the present application is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application.
  • the aforementioned readable storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc, etc., which can store program codes. medium.
  • the application after the application obtains the feature map to be processed of the target convolution kernel of the feature extraction convolutional neural network, it will Cut the input feature map obtained by image preprocessing from the feature map to be processed into multiple input feature submaps according to the preset image size, and fill the boundary of each input feature submap to obtain the corresponding feature submap to be convoluted , and then for each feature subgraph to be convoluted, use the Karatsuba algorithm to reduce the multiplier resource loss during the convolution operation between the target convolution kernel and the feature subgraph to be convolved, and effectively suppress the addition during the convolution operation
  • the corresponding target feature submap is obtained, and then according to the distribution of each input feature submap in the input feature map, the boundary superposition of the target feature submaps corresponding to the adjacent positions is performed to obtain the overlapping feature map, and finally the The overlapping feature map is circumscribed to obtain the corresponding output

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)

Abstract

本申请提供一种图像特征提取方法及装置、计算机设备和可读存储介质，涉及卷积神经网络技术领域。本申请在获取到特征提取卷积神经网络的目标卷积核的待处理特征图后，会按照预设图像尺寸将由该待处理特征图经图像预处理得到的输入特征图切割为多个输入特征子图，并对每个输入特征子图进行边界填充，得到待卷积特征子图，而后针对每个待卷积特征子图，通过Karatsuba算法消减目标卷积核与该待卷积特征子图的卷积运算时的乘法器资源损耗，并有效抑制卷积运算时的加法器资源损耗的增长，得到目标特征子图，接着将多个目标特征子图进行边界叠加及边界环切，得到输出特征图，从而从整体上降低特征提取时的计算资源损耗，提升特征提取效率。

Description

图像特征提取方法及装置、计算机设备和可读存储介质
相关申请的交叉引用
本申请要求于2021年08月11日提交中国国家知识产权局的申请号为202110917990.5、名称为“图像特征提取方法及装置、计算机设备和可读存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及卷积神经网络技术领域，具体而言，涉及一种图像特征提取方法及装置、计算机设备和可读存储介质。
背景技术
随着科学技术的不断发展，机器学习技术在各大行业的应用越发广阔，作为近年来深度学习研究和应用领域中最活跃的一个分支——卷积神经网络技术，更是随着人工智能技术的不断推进被人们注入了更多的关注度，人们对于卷积神经网络技术的具体应用提出了更多的诉求，例如要求卷积神经网络在实际应用(例如，人脸特征提取、图像特征处理等操作)时具有更小的计算资源损耗以及更高的计算速率。
发明内容
有鉴于此，本申请的目的在于提供一种图像特征提取方法及装置、计算机设备和可读存储介质，能够降低特征提取时的计算资源损耗，并提升特征提取效率。
为了实现上述目的，本申请实施例采用的技术方案如下：
第一方面，本申请提供一种图像特征提取方法，所述方法包括：
获取针对特征提取卷积神经网络的目标卷积核的待处理特征图；
根据预设图像尺寸对所述待处理特征图进行图像预处理，得到对应的输入特征图，其中所述输入特征图的图像尺寸为所述预设图像尺寸的整数倍；
按照所述预设图像尺寸将所述输入特征图切割为多个输入特征子图；
针对每个所述输入特征子图，根据所述目标卷积核的卷积核尺寸对该输入特征子图进行边界填充，得到对应的待卷积特征子图；
针对每个所述待卷积特征子图，基于Karatsuba算法采用所述目标卷积核对该待卷积特征子图进行卷积运算，得到对应的目标特征子图；
针对得到的多个所述目标特征子图，根据所述输入特征图中各个所述输入特征子图的分布状况，将对应位置相邻的所述目标特征子图进行边界叠加，得到重叠特征图；
根据所述目标卷积核的卷积核尺寸对所述重叠特征图进行边界环切，得到对应的输出特征图，其中所述输出特征图用于表示与所述待处理特征图匹配的所述输入特征图的特征提取结果。
在可选的实施方式中，所述针对每个所述输入特征子图，根据所述目标卷积核的卷积核尺寸对该输入特征子图进行边界填充，得到对应的待卷积特征子图的步骤，包括：
根据所述目标卷积核的卷积核尺寸，计算单个输入特征子图所需空白元素环的第一目标数目，其中所述第一目标数目由所述卷积核尺寸所对应的行/列数减一得到；
针对每个所述输入特征子图，以该输入特征子图为中心在该输入特征子图周围外扩地依次填充第一目标数目个空白元素环，得到对应的待卷积特征子图。
在可选的实施方式中，所述针对每个所述待卷积特征子图，基于Karatsuba算法采用所述目标卷积核对该待卷积特征子图进行卷积运算，得到对应的目标特征子图的步骤，包括：
针对每个所述待卷积特征子图，根据对应的输入特征子图中不同分布位置的特征元素值，基于Karatsuba算法构建该输入特征子图的行列特征表达式；
针对所述目标卷积核，根据该目标卷积核中不同分布位置的卷积元素值，基于Karatsuba算法构建该目标卷积核的行列特征表达式；
针对每个所述待卷积特征子图，基于Karatsuba算法将对应输入特征子图的行列特征表达式与所述目标卷积核的行列特征表达式的乘法运算过程进行拆分换算，得到对应的拆分换算结果；
针对每个所述待卷积特征子图，将对应的拆分换算结果基于与目标卷积核对应的预设卷积映射分布状况进行元素表达，得到对应的目标特征子图。
在可选的实施方式中，所述根据所述输入特征图中各个所述输入特征子图的分布状况，将对应位置相邻的所述目标特征子图进行边界叠加，得到重叠特征图的步骤，包括：
根据所述目标卷积核的卷积核尺寸，确定单个目标特征子图的待叠加元素环的第二目标数目，其中所述第二目标数目由所述卷积核尺寸所对应的行/列数减一得到；
针对每个目标特征子图，由外往内地选取该目标特征子图的第二目标数目个子图元素环作为待叠加元素环，得到该目标特征子图的待叠加区域；
按照所述输入特征图中各个所述输入特征子图的分布状况，对各目标特征子图进行图像拼接，得到对应的图像拼接结果；
在所述图像拼接结果中将位置相邻的待叠加区域各自的接触边缘区域进行元素叠加，得到所述重叠特征图。
在可选的实施方式中，所述根据所述目标卷积核的卷积核尺寸对所述重叠特征图进行边界环切，得到对应的输出特征图的步骤，包括：
根据所述目标卷积核的卷积核尺寸，确定所述重叠特征图的待切除元素环的第三目标数目，其中所述第三目标数目由所述卷积核尺寸所对应的行/列数减一得到；
由外往内地选取该重叠特征图的第三目标数目个特征图元素环作为待切除元素环，得到该重叠特征图的待切除区域；
对该重叠特征图中的待切除区域进行图像切除，得到与所述输入特征图对应的所述输出特征图。
在可选的实施方式中，所述方法还包括：
根据所述目标卷积核的卷积核尺寸对所述预设图像尺寸进行配置，其中所述预设图像尺寸所对应的行数大于或等于所述卷积核尺寸所对应的行数，所述预设图像尺寸所对应的列数大于或等于所述卷积核尺寸所对应的列数。
第二方面，本申请提供一种图像特征提取装置，所述装置包括：
特征图获取模块，用于获取针对特征提取卷积神经网络的目标卷积核的待处理特征图；
特征图预处理模块，用于根据预设图像尺寸对所述待处理特征图进行图像预处理，得到对应的输入特征图，其中所述输入特征图的图像尺寸为所述预设图像尺寸的整数倍；
输入特征切割模块，用于按照预设图像尺寸将所述输入特征图切割为多个输入特征子图，其中每个所述输入特征子图的图像尺寸与所述预设图像尺寸保持一致；
子图边界填充模块，用于针对每个所述输入特征子图，根据所述目标卷积核的卷积核尺寸对该输入特征子图进行边界填充，得到对应的待卷积特征子图；
特征卷积运算模块，用于针对每个所述待卷积特征子图，基于Karatsuba算法采用所述目标卷积核对该待卷积特征子图进行卷积运算，得到对应的目标特征子图；
子图边界叠加模块，用于针对得到的多个所述目标特征子图，根据所述输入特征图中各个所述输入特征子图的分布状况，将对应位置相邻的所述目标特征子图进行边界叠加，得到重叠特征图；
输出特征提取模块，用于根据所述目标卷积核的卷积核尺寸对所述重叠特征图进行边界环切，得到对应的输出特征图，其中所述输出特征图用于表示与所述待处理特征图匹配的所述输入特征图的特征提取结果。
在可选的实施方式中，所述装置还包括：
切割尺寸配置模块，用于根据所述目标卷积核的卷积核尺寸对所述预设图像尺寸进行配置，其中所述预设图像尺寸所对应的行数大于或等于所述卷积核尺寸所对应的行数，所述预设图像尺寸所对应的列数大于或等于所述卷积核尺寸所对应的列数。
第三方面，本申请提供一种计算机设备，包括处理器和存储器，所述存储器存储有能够被所述处理器执行的计算机程序，所述处理器可执行所述计算机程序，实现前述实施方式中任意一项所述的图像特征提取方法。
第四方面，本申请提供一种可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时，实现前述实施方式中任意一项所述的图像特征提取方法。
在此情况下，本申请实施例的有益效果包括以下内容：
本申请在获取到特征提取卷积神经网络的目标卷积核的待处理特征图后，会按照预设图像尺寸将由该待处理特征图经图像预处理得到的输入特征图切割为多个输入特征子图，并对每个输入特征子图进行边界填充，得到对应的待卷积特征子图，而后针对每个待卷积特征子图，通过Karatsuba算法消减目标卷积核与该待卷积特征子图的卷积运算过程中的乘法器资源损耗，并有效抑制卷积运算过程中的加法器 资源损耗的增长，得到对应的目标特征子图，接着根据输入特征图中各输入特征子图的分布状况，将对应位置相邻的目标特征子图进行边界叠加，得到重叠特征图，最后对该重叠特征图进行边界环切，得到对应的输出特征图，从而从整体上降低特征提取时的计算资源损耗，并提升特征提取效率。
为使本申请的上述目的、特征和优点能更明显易懂，下文特举较佳实施例，并配合所附附图，作详细说明如下。
附图说明
为了更清楚地说明本申请实施例的技术方案，下面将对实施例中所需要使用的附图作简单地介绍，应当理解，以下附图仅示出了本申请的某些实施例，因此不应被看作是对范围的限定，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他相关的附图。
图1为本申请实施例提供的计算机设备的组成示意图；
图2为本申请实施例提供的图像特征提取方法的流程示意图之一；
图3为图2中的步骤S240包括的子步骤的流程示意图；
图4为本申请实施例提供的待卷积特征子图的分布示意图；
图5为图2中的步骤S250包括的子步骤的流程示意图；
图6为本申请实施例提供的预设卷积映射分布状况的展示示意图之一；
图7为本申请实施例的预设卷积映射分布状况的展示示意图之二；
图8为图2中的步骤S260包括的子步骤的流程示意图；
图9为本申请实施例提供的图像拼接结果的展示示意图；
图10为图2中的步骤S270包括的子步骤的流程示意图；
图11为本申请实施例提供的重叠特征图的分布示意图；
图12为本申请实施例提供的图像特征提取方法的流程示意图之二；
图13为本申请实施例提供的图像特征提取装置的组成示意图之一；
图14为本申请实施例提供的图像特征提取装置的组成示意图之二。
图标：10-计算机设备；11-存储器；12-处理器；13-通信单元；100-图像特征提取装置；110-特征图获取模块；120-特征图预处理模块；130-输入特征切割模块；140-子图边界填充模块；150-特征卷积运算模块；160-子图边界叠加模块；170-输出特征提取模块；180-切割尺寸配置模块。
具体实施方式
为使本申请实施例的目的、技术方案和优点更加清楚，下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。通常在此处附图中描述和示出的本申请实施例的组件可以以各种不同的配置来布置和设计。
因此，以下对在附图中提供的本申请的实施例的详细描述并非旨在限制要求保护的本申请的范围， 而是仅仅表示本申请的选定实施例。基于本申请中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
应注意到：相似的标号和字母在下面的附图中表示类似项，因此，一旦某一项在一个附图中被定义，则在随后的附图中不需要对其进行进一步定义和解释。
在本申请的描述中，需要理解的是，术语“第一”和“第二”等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。对于本领域的普通技术人员而言，可以具体情况理解上述术语在本申请中的具体含义。
申请人通过辛苦调研发现，在卷积神经网络的具体应用过程中可采用快速傅里叶变换(FFT)或Winograd算法对卷积运算进行优化，以降低卷积神经网络的原本所需计算资源损耗，从而提升运算效率。
但需要注意的是，采用FFT算法实现的卷积优化过程是通过将时域内的卷积运算转化到评语内的乘积运算，其中涉及到复数的运算，一次复数乘法相当于四次实数乘加操作，极大地增大卷积神经网络运行时的内存带宽需求，从整体来看乘法器资源损耗及加法器资源损耗仍然较多，卷积优化效果不佳。
而采用Winograd算法实现的卷积优化过程是通过将数据映射到另一个实数空间上，采用加减运算代替部分乘法运算，从而在“加减运算速度远远高于乘法运算，且乘法器运行时所需计算资源损耗远远高于加法器运行时所需计算资源损耗”的情况下提升运算效率，其中往往需要大量加法器资源损耗才能完成对部分乘法运算的替代操作，故而该方式实质存在加法器资源损耗过大的问题，整体的卷积优化效果仍然显得不佳。
在此情况下，为有效降低特征提取卷积神经网络在执行特征提取操作时的计算资源损耗，并提升特征提取效率，即提高该特征提取卷积神经网络的卷积优化效果，本申请实施例通过提供一种图像特征提取方法及装置、计算机设备和可读存储介质实现前述功能。
下面结合附图，对本申请的一些实施方式作详细说明。在不冲突的情况下，下述的实施例及实施例中的特征可以相互结合。
请参照图1，图1是本申请实施例提供的计算机设备10的组成示意图。在本申请实施例中，所述计算机设备10用于快速实现图像特征提取操作，其中所述计算机设备10存储有特征提取卷积神经网络，并可在基于该特征提取卷积神经网络对待处理图像进行特征提取的过程中，对该特征提取卷积神经网络的卷积运算过程进行有效优化，使原始卷积运算过程所对应的乘法器资源损耗得到有效消减，并同步地有效抑制原始卷积运算过程所对应的加法器资源损耗的增长，从而从整体上降低特征提取时的计算资源损耗，并提升特征提取效率，提高了该神经网络的卷积优化效果。其中，所述计算机设备10可以是， 但不限于，智能手机、平板电脑、个人计算机、服务器等。
在本实施例中，所述计算机设备10可以包括存储器11、处理器12、通信单元13及图像特征提取装置100。所述存储器11、所述处理器12及所述通信单元13各个元件相互之间直接或间接地电性连接，以实现数据的传输或交互。例如，所述存储器11、所述处理器12及所述通信单元13这些元件相互之间可通过一条或多条通讯总线或信号线实现电性连接。
在本实施例中，所述存储器11可以是，但不限于，随机存取存储器(Random Access Memory，RAM)，只读存储器(Read Only Memory，ROM)，可编程只读存储器(Programmable Read-Only Memory，PROM)，可擦除只读存储器(Erasable Programmable Read-Only Memory，EPROM)，电可擦除只读存储器(Electric Erasable Programmable Read-Only Memory，EEPROM)等。其中，所述存储器11用于存储计算机程序，所述处理器12在接收到执行指令后，可相应地执行所述计算机程序。所述存储器11还用于存储特征提取卷积神经网络，以便于使用该特征提取卷积神经网络实现图像特征提取操作，其中所述特征提取卷积神经网络可以包括至少一个卷积层，每个卷积层可以包括至少一个用于提取图像特征的特殊卷积核，所述特殊卷积核可以是3*3的卷积核，也可以是5*5卷积核。
在本实施例中，所述处理器12可以是一种具有信号的处理能力的集成电路芯片。所述处理器12可以是通用处理器，包括中央处理器(Central Processing Unit，CPU)、图形处理器(Graphics Processing Unit，GPU)及网络处理器(Network Processor，NP)、数字信号处理器(DSP)、专用集成电路(ASIC)、现成可编程门阵列(FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件中的至少一种。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等，可以实现或者执行本申请实施例中的公开的各方法、步骤及逻辑框图。
在本实施例中，所述通信单元13用于通过网络建立所述计算机设备10与其他电子设备之间的通信连接，并通过所述网络收发数据，其中所述网络包括有线通信网络及无线通信网络。例如，所述计算机设备10可通过所述通信单元13与智能手机通信连接，以从所述智能手机处获取需要特征提取的待处理图像。
在本实施例中，所述图像特征提取装置100包括至少一个能够以软件或固件的形式存储于所述存储器11中或者在所述计算机设备10的操作系统中的软件功能模块。所述处理器12可用于执行所述存储器11存储的可执行模块，例如所述图像特征提取装置100所包括的软件功能模块及计算机程序等。所述计算机设备10可通过所述图像特征提取装置100在使用特征提取卷积神经网络对待处理图像进行特征提取的过程中，对该特征提取卷积神经网络的卷积运算过程进行有效优化，使原始卷积运算过程所对应的乘法器资源损耗得到有效消减，并同步地有效抑制原始卷积运算过程所对应的加法器资源损耗的增长，从而从整体上降低特征提取时的计算资源损耗，并提升特征提取效率，提高该神经网络的卷积优化效果。
可以理解的是，图1所示的框图仅为所述计算机设备10的一种组成示意图，所述计算机设备10还 可包括比图1中所示更多或者更少的组件，或者具有与图1所示不同的配置。图1中所示的各组件可以采用硬件、软件或其组合实现。
在本申请中，为确保所述计算机设备10能够有效消减图像特征提取过程中的乘法器资源损耗，并同步地抑制图像特征提取过程中的加法器资源损耗的增长，提升神经网络的卷积优化效果，本申请实施例通过提供一种图像特征提取方法实现前述功能，下面对本申请提供的图像特征提取方法进行详细阐述。
请参照图2，图2是本申请实施例提供的图像特征提取方法的流程示意图之一。在本申请实施例中，图2所示的图像特征提取方法可以包括步骤S210～步骤S260。
步骤S210，获取针对特征提取卷积神经网络的目标卷积核的待处理特征图。
在本实施例中，所述目标卷积核为该特征提取卷积神经网络中的某个需要实现特征提取操作的卷积核，该目标卷积核的卷积核尺寸可以是，但不限于，3*3、5*5等。所述待处理特征图即为需要采用该目标卷积核实现特征提取操作的特征图像。
步骤S220，根据预设图像尺寸对待处理特征图进行图像预处理，得到对应的输入特征图。
在本实施例中，所述预设图像尺寸用于表征所述输入特征图被切割时的标准尺寸，因此当所述计算机设备10得到待处理特征图后，会基于所述预设图像尺寸对该待处理特征图进行图像预处理，使图像预处理得到的所述输入特征图所具有的图像尺寸为所述预设图像尺寸的整数倍。其中，可以理解的是，若所述待处理特征图的图像尺寸并非所述预设图像尺寸的整数倍，则所述图像预处理将对应包括图像填充操作，可在所述待处理特征图的边界周围或某一侧边缘位置处填充元素值为零的空白元素，使最终得到输入特征图的图像尺寸为所述预设图像尺寸的整数倍。在本实施例的一种实施方式中，所述图像预处理还可以包括图像数字化处理、图像几何变换处理、图像归一化处理、图像平滑处理、图像复原处理及图像增强等操作中的任意一种或多种组合。
步骤S230，按照预设图像尺寸将输入特征图切割为多个输入特征子图。
在本实施例中，所述预设图像尺寸用于表征所述输入特征图被切割时的标准尺寸，每个所述输入特征子图的子图尺寸与该预设图像尺寸保持一致。其中，所述预设图像尺寸通常需要等于或大于所述目标卷积核的卷积核尺寸，即所述预设图像尺寸所对应的图像元素行数大于或等于所述目标卷积核的卷积核尺寸所对应的卷积元素行数，所述预设图像尺寸所对应的图像元素列数大于或等于所述目标卷积核的卷积核尺寸所对应的卷积元素列数。切割出的多个所述输入特征子图相互拼接即可得到所述输入特征图。
步骤S240，针对每个输入特征子图，根据目标卷积核的卷积核尺寸对该输入特征子图进行边界填充，得到对应的待卷积特征子图。
在本实施例中，所述待卷积特征子图用于表示可直接与所述目标卷积核进行卷积运算的特征子图。所述计算机设备10在得到输入特征图所对应的多个输入特征子图后，可根据对应目标卷积核的卷积核尺寸，对单个输入特征子图进行边界填充，以确保得到的待卷积特征子图与该目标卷积核在卷积运算维度内相互适配。
可选地，请结合参照图3及图4，其中图3是图2中的步骤S240包括的子步骤的流程示意图，图4是本申请实施例提供的待卷积特征子图的分布示意图。在本实施例中，所述步骤S240可以包括子步骤S241～子步骤S242。
子步骤S241，根据目标卷积核的卷积核尺寸，计算单个输入特征子图所需空白元素环的第一目标数目，其中第一目标数目由卷积核尺寸所对应的行/列数减一得到。
其中，所述空白元素环用于表征该空白元素环所包括的多个空白元素相互紧靠且呈中空环状分布，所述空白元素的元素值为零。所述第一目标数目用于表示对应输入特征子图在实现边界填充时所需填充的空白元素环个数，所述第一目标数目由所述目标卷积核的卷积核尺寸所对应的行/列数减一得到。
子步骤S242，针对每个输入特征子图，以该输入特征子图为中心在该输入特征子图周围外扩地依次填充第一目标数目个空白元素环，得到对应的待卷积特征子图。
在本实施例中，所述计算机设备10在确定单个输入特征子图所需填充的空白元素环个数(即所述第一目标数目)后，会相应地在对应输入特征子图周围边界套设式地填充所述第一目标数目个空白元素环，以形成对应的待卷积特征子图，其中若填充有多个空白元素环，则相邻两个空白元素环中元素数目较少的空白元素环嵌设在元素数目较多的空白元素环的中空部位内。以图4所示的待卷积特征子图为例，若目标卷积核的卷积核尺寸为3*3，且输入特征子图(由图4中的多个实心格子拼接形成，每个实心格子代表一个输入特征子图的子图元素)的子图尺寸也为3*3，则针对该输入特征子图的第一目标数目为3-1＝2，而后将以该输入特征子图为中心在该输入特征子图的边界周围外扩地依次填充2个空白元素环，使这2个空白元素环相互嵌套在一起，越靠近该输入特征子图的空白元素环的元素数目越少。
由此，本申请可通过执行上述子步骤S241及子步骤S242，确保构建得到的待卷积特征子图与对应的目标卷积核在卷积运算维度内相互适配。
步骤S250，针对每个待卷积特征子图，基于Karatsuba算法采用目标卷积核对该待卷积特征子图进行卷积运算，得到对应的目标特征子图。
在本实施例中，所述计算机设备10在某个输入特征子图所对应的待卷积特征子图后，可在该待卷积特征子图与目标卷积核进行卷积运算的过程中运用Karatsuba算法进行相关处理，得到完成卷积运算操作的与所述输入特征子图对应的输入特征子图，从而通过Karatsuba算法使用少量加法器资源损耗来代替卷积运算过程中的部分原有乘法计算，以消减目标卷积核与该待卷积特征子图的卷积运算过程中的乘法器资源损耗，并有效抑制卷积运算过程中的加法器资源损耗的增长。其中，Karatsuba算法能够基于分治思想将参与运算的乘数与被乘数的乘积运算拆分成三个部分积的运算，从而消减原乘积操作所需的乘法计算资源损耗，并能使用少量加法计算资源损耗实现相应地乘法计算效果。
可选地，请参照图5、图6及图7，其中图5是图2中的步骤S250包括的子步骤的流程示意图，图6为本申请实施例提供的预设卷积映射分布状况的展示示意图之一，图7为本申请实施例的预设卷积映射分布状况的展示示意图之二。在本实施例中，所述步骤S250可以包括子步骤S251～子步骤S254。
子步骤S251，针对每个所述待卷积特征子图，根据对应的输入特征子图中不同分布位置的特征元素值，基于Karatsuba算法构建该输入特征子图的行列特征表达式。
在本实施例中，所述待卷积特征子图的行列特征表达式用于表征对应待卷积特征子图中与输入特征子图匹配的有效图像特征。在基于Karatsuba算法构建匹配的行列特征表达式过程中，可将对应输入特征子图的有效图像特征视为不同分布位置的特征元素值按照行列分布综合形成的数值，此时可基于该输入特征子图中处于同一元素行的各特征元素值，运用Karatsuba算法构建出该元素行的行特征表达式，所述行特征表达式以该元素行的各特征元素值作为元素行表达式权重，并以各特征元素值的分布位数为行特征表达幂，采用行特征表达基进行乘加组成得到。而后，使用不同元素行的行特征表达式作为该输入特征子图的元素列表达式权重，并以各元素行的行特征表达式的分布位数为列特征表达幂，采用列特征表达基进行乘加组成得到所述输入特征子图的行列特征表达式。
以图6所示的预设卷积映射分布状况的展示示意图为例，针对组成与所述待卷积特征子图类似的一维参考特征图来说，该一维参考特征图所对应的有效特征子图内容为非空白元素a1、a2及a3，这三个非空白元素处于同一元素行，则该一维参考特征图的行列特征表达式即可为A＝a1+a2*X n+a3*X 2n，其中X用于表示所述行特征表达基，n用于表示各非空白元素的分布位数。
以图7所示的预设卷积映射分布状况的展示示意图为例，针对组成与所述待卷积特征子图类似的二维参考特征图来说，该二维参考特征图所对应的有效特征子图内容为非空白元素a1、a2、a3、a4、a5、a6、a7、a8及a9，前三个非空白元素处于同一元素行，中三个非空白元素处于同一元素行，后三个非空白元素处于同一元素行，则该二维参考特征图将对应存在3个元素行的行特征表达式，即为A 1＝a1+a2*X n+a3*X 2n，A 2＝a4+a5*X n+a6*X 2n及A 3＝a7+a8*X n+a9*X 2n，接着该二维参考特征图的行列特征表达式即为A＝A 1+A 2*Y m+A 3*Y 2m，其中Y用于表示所述列特征表达基，m用于表示各元素行的分布位数。
子步骤S252，针对目标卷积核，根据该目标卷积核中不同分布位置的卷积元素值，基于Karatsuba算法构建该目标卷积核的行列特征表达式。
在本实施例中，所述目标卷积核的行列特征表达式用于表示对应目标卷积核的有效图像特征。在基于Karatsuba算法构建匹配的行列特征表达式过程中，可将对应目标卷积核的有效图像特征视为不同分布位置的卷积元素值按照行列分布综合形成的数值，此时可基于该目标卷积核中处于同一元素行的各卷积元素值，运用Karatsuba算法构建出该元素行的行特征表达式，所述行特征表达式以该元素行的各卷积元素值作为元素行表达式权重，并以各卷积元素值的分布位数为行特征表达幂，采用行特征表达基进行乘加组成得到。而后，使用不同元素行的行特征表达式作为该目标卷积核的元素列表达式权重，并以各元素行的行特征表达式的分布位数为列特征表达幂，采用列特征表达基进行乘加组成得到所述目标卷积核的行列特征表达式。
以图6所示的预设卷积映射分布状况的展示示意图为例，针对与目标卷积核组成类似的一维参考卷积核来说，该一维参考卷积核所对应的有效特征为非空白元素b1、b2及n3，这三个非空白元素处于同一元素行，则该一维参考卷积核的行列特征表达式即可为B＝b1+b2*X n+b3*X 2n。
以图7所示的预设卷积映射分布状况的展示示意图为例，针对与目标卷积核组成类似的二维参考卷积核来说，该二维参考卷积核所对应的有效特征为非空白元素b1、b2、b3、b4、b5、b6、b7、b8及b9，前三个非空白元素处于同一元素行，中三个非空白元素处于同一元素行，后三个非空白元素处于同一元素行，则该二维参考卷积核将对应存在3个元素行的行特征表达式，即为B 1＝b1+b2*X n+b3*X 2n，B 2＝b4+b5*X n+b6*X 2n及B 3＝b7+b8*X n+b9*X 2n，接着该二维参考卷积核的行列特征表达式即为B＝B 1+B 2*Y m+B 3*Y 2m。
子步骤S253，针对每个待卷积特征子图，基于Karatsuba算法将对应输入特征子图的行列特征表达式与目标卷积核的行列特征表达式的乘法运算过程进行拆分换算，得到对应的拆分换算结果。
在本实施例中，当确定某个待卷积特征子图所对应的行列特征表达式后，可通过将该待卷积特征子图的行列特征表达式与所述目标卷积核的行列特征表达式之间的乘法运算表达式中的各项权重系数进行拆分换算，以通过使用少量加法计算资源损耗实现相应地乘法计算效果，消减原乘积操作所需的乘法计算资源损耗，得到对应的拆分换算结果。
以图6所示的预设卷积映射分布状况的展示示意图为例，若一维参考特征图与一维参考卷积核各自的行列特征表达式进行乘法运算，则对应的乘法运算过程可表示为A×B＝a3·b3*X 4n+(a3·b2+a2·b3)*X 3n+(a3·b1+a2·b2+a1·b3)*X 2n+(a1·b2+a2·b1)*X n+a1·b1，而后采用
Figure PCTCN2022078874-appb-000001
进行拆分换算，得到对应的拆分换算结果为
Figure PCTCN2022078874-appb-000002
以图7所示的预设卷积映射分布状况的展示示意图为例，若二维参考特征图与二维参考卷积核各自的行列特征表达式进行乘法运算，则对应的乘法运算过程可表示为A×B＝A 3×B 3*Y 4m+(A 3×B 2+A 2×B 3)*Y 3m+(A 3·B 1+A 2×B 2+A 2×B 3)*Y 2m+(A 1×B 2+A 2×B 1)*Y m+A 1×B 1，而后采用如
Figure PCTCN2022078874-appb-000003
类似的拆分公式组对该乘法运算 过进行拆分换算，得到对应的拆分换算结果为
Figure PCTCN2022078874-appb-000004
由此，本申请可通过执行上述子步骤S253，仅使用少量加法计算资源损耗实现相应地乘法计算效果，消减原乘积操作所需的乘法计算资源损耗，得到对应的拆分换算结果。
子步骤S254，针对每个待卷积特征子图，将对应的拆分换算结果基于与目标卷积核对应的预设卷积映射分布状况进行元素表达，得到对应的目标特征子图。
在本实施例中，所述预设卷积映射分布状况用于表示在基于相互适配的参考特征图与参考卷积核通过相关卷积运算得到的特征图结果中，该特征图结果中不同位置的特征元素值与参考特征图中各特征元素值和参考卷积核中各卷积元素值之间的计算映射关系，例如图6中的一维预设卷积映射分布状况与一维参考特征图及一维参考卷积核相互适配，图7中的二维预设卷积映射分布状况与二维参考特征图及二维参考卷积核相互适配。所述计算机设备10在得到某个待卷积特征子图所对应的与目标卷积核匹配的拆分换算结果后，可采用格式与该待卷积特征子图及目标卷积核匹配的预设卷积映射分布状况对该拆分换算结果进行特征图构建，并对构建出的特征图中各元素的具体元素值进行表达，得到与该待卷积特征子图对应的目标特征子图。
由此，本申请可通过执行上述子步骤S251～子步骤S254，使用少量加法器资源损耗来代替卷积运算过程中的部分原有乘法计算，以消减目标卷积核与该待卷积特征子图的卷积运算过程中的乘法器资源损耗，并有效抑制卷积运算过程中的加法器资源损耗的增长。
步骤S260，针对得到的多个目标特征子图，根据输入特征图中各个输入特征子图的分布状况，将对应位置相邻的目标特征子图进行边界叠加，得到重叠特征图。
在本实施例中，当确定出每个输入特征子图分别对应的经目标卷积核卷积处理后的目标特征子图后，可通过将这多个目标特征子图进行边界叠加，以实现对多个目标特征子图的特征拼接操作。
可选地，请参照图8及图9，其中图8是图2中的步骤S260包括的子步骤的流程示意图，图9是本申请实施例提供的图像拼接结果的展示示意图。在本实施例中，所述步骤S260可以包括子步骤S261～子步骤S264。
子步骤S261，根据目标卷积核的卷积核尺寸，确定单个目标特征子图的待叠加元素环的第二目标数目，其中第二目标数目由卷积核尺寸所对应的行/列数减一得到。
在本实施例中，所述待叠加元素环用于表示对应目标特征子图中呈中空环形分布的需要与其他目标特征子图进行元素叠加的多个特征元素。所述第二目标数目用于表示对应目标特征子图的靠近边界的待叠加元素环个数，其中同一目标特征子图的相邻待叠加元素环中远离边界的待叠加元素环被靠近边界的待叠加元素环嵌套，所述第二目标数目由所述目标卷积核的卷积核尺寸所对应的行/列数减一得到。
子步骤S262，针对每个目标特征子图，由外往内地选取该目标特征子图的第二目标数目个子图元素 环作为待叠加元素环，得到该目标特征子图的待叠加区域。
在本实施例中，所述计算机设备10在确定单个目标特征子图所需叠加的待叠加元素环个数(即第二目标数目)后，会相应地从该目标特征子图边界开始由外往内地选取该目标特征子图的第二目标数目个子图元素环作为待叠加元素环，并将选取的所有待叠加元素环组合形成该目标特征子图的待叠加区域。其中，所述子图元素环用于表示对应目标特征子图中呈中空环形分布的多个相互紧靠的子图元素。以图9所示的图像拼接结果的展示示意图为例，若目标卷积核的卷积核尺寸为3*3，则在目标特征子图中，对应的第二目标数目为3-1＝2，而后将从该目标特征子图边界开始由外往内地选取2个相互嵌套的子图元素环作为所述待叠加区域，使该目标特征子图中剩余子图元素构成该目标特征子图的非叠加区域(由图9中相互紧靠的多个实心格子拼接形成，每个实心格子代表一个目标特征子图的非叠加区域中的子图元素)。
子步骤S263，按照输入特征图中各个输入特征子图的分布状况，对各目标特征子图进行图像拼接，得到对应的图像拼接结果。
在本实施例中，多个目标特征子图在相互拼接时，会按照该目标特征子图所对应的输入特征子图在所述输入特征图中的分布位置进行排列，从而使得对应得到的图像拼接结果在特征位置分布上与所述输入特征图类似，如图9中的4个目标特征子图相互拼接所得到的图像拼接结果。
子步骤S264，在图像拼接结果中将位置相邻的待叠加区域各自的接触边缘区域进行元素叠加，得到重叠特征图。
在本实施例中，当不同目标特征子图各自的待叠加区域在图像拼接结果中排列，相邻的待叠加区域相互接触时，每个待叠加区域上的接触边缘区域即为从该待叠加区域的与其他待叠加区域接触的边缘位置到所在目标特征子图中非叠加区域之间的区域范围，如在位于图9上方的两个相邻目标特征子图中，左上方的目标特征子图所具有的接触边缘区域为对应待叠加区域中被纵向虚线框所圈定的区域范围，右上方的目标特征子图所具有的接触边缘区域为对应待叠加区域中被纵向虚线框所圈定的区域范围，这两个目标特征子图各自的接触边缘区域相互接触。而本申请通过将图像拼接结果中位置相邻的待叠加区域各自的接触边缘区域进行元素叠加，实现对该图像拼接结果中多个目标特征子图各自的子图特征进行特征拼接操作，得到包括有干扰噪声的能够表征出所述输入特征图的特征提取结果的重叠特征图。
由此，本申请可通过执行上述子步骤S261～子步骤S264，实现对多个目标特征子图的特征拼接操作，得到包括有干扰噪声的能够表征出所述输入特征图的特征提取结果的重叠特征图。
步骤S270，根据目标卷积核的卷积核尺寸对重叠特征图进行边界环切，得到对应的输出特征图。
在本实施例中，所述输出特征图用于表示与所述待处理特征图匹配的所述输入特征图的特征提取结果，所述边界环切操作可去除所述重叠特征图中的干扰噪声，保留最终的与该输入特征图匹配的特征提取结果。
可选地，请参照图10及图11，其中图10是图2中的步骤S270包括的子步骤的流程示意图，图11 是本申请实施例提供的重叠特征图的分布示意图。在本实施例中，所述步骤S270可以包括子步骤S271～子步骤S273。
子步骤S271，根据目标卷积核的卷积核尺寸，确定重叠特征图的待切除元素环的第三目标数目，其中第三目标数目由卷积核尺寸所对应的行/列数减一得到。
在本实施例中，所述待环切元素环用于表示所述重叠特征图中呈中空环形分布的需要去除的多个特征元素，所述第三目标数目用于表示对应重叠特征图的靠近边界的待环切元素环个数，其中同一重叠特征图的相邻待环切元素环中远离边界的待环切元素环被靠近边界的待环切元素环嵌套，所述第三目标数目由所述目标卷积核的卷积核尺寸所对应的行/列数减一得到。
子步骤S272，由外往内地选取该重叠特征图的第三目标数目个特征图元素环作为待切除元素环，得到该重叠特征图的待切除区域。
在本实施例中，所述计算机设备10在确定所述重叠特征图所对应的第三目标数目后，会相应地从该重叠特征图边界开始由外往内地选取该重叠特征图的第三目标数目个特征图元素环作为待切除元素环，并将选取的所有待切除元素环组合形成该重叠特征图的待切除区域。其中，所述特征图元素环用于表示对应重叠特征图中呈中空环形分布的多个相互紧靠的特征图元素。以图11所示的重叠特征图为例，若目标卷积核的卷积核尺寸为3*3，则在重叠特征图中对应的第三目标数目为3-1＝2，而后将从该重叠特征图边界开始由外往内地选取2个相互嵌套的特征图元素环作为所述待切除区域。
子步骤S273，对该重叠特征图中的待切除区域进行图像切除，得到与输入特征图对应的输出特征图。
由此，本申请可通过执行上述子步骤S271～子步骤S273，去除所述重叠特征图中的干扰噪声，保留最终的与该输入特征图匹配的特征提取结果。
在此情况下，本申请得以通过执行上述步骤S210～步骤S270，在图像特征提取过程中消减卷积运算过程所需的乘法器资源损耗，并有效抑制卷积运算过程所需的加法器资源损耗的增长，从而从整体上降低特征提取时的计算资源损耗，并提升特征提取效率，提升特征提取卷积神经网络的卷积优化效果。
可选地，请参照图12，图12是本申请实施例提供的图像特征提取方法的流程示意图之二。在本申请实施例中，图12所示的图像特征提取方法与图2所示的图像特征提取方法相比，图12所示的图像特征提取方法还可以包括步骤S209。
步骤S209，根据目标卷积核的卷积核尺寸对预设图像尺寸进行配置。
在本实施例中，配置出的所述预设图像尺寸所对应的行数图像元素行数大于或等于所述目标卷积核的卷积核尺寸所对应的卷积元素行数，配置出的所述预设图像尺寸所对应的图像元素列数大于或等于所述目标卷积核的卷积核尺寸所对应的卷积元素列数。
由此，本申请可通过执行所述步骤S209，确保切割出的多个输入特征子图相互拼接即可得到与待处理特征图匹配的输入特征图，并可采用Karatsuba算法使对应输入特征子图与目标卷积核进行卷积运算。
在本申请中，为确保所述计算机设备10能够通过所述图像特征提取装置100执行上述图像特征提 取方法，本申请通过对所述图像特征提取装置100进行功能模块划分的方式实现前述功能。下面对本申请提供的图像特征提取装置100的具体组成进行相应描述。
请参照图13，图13是本申请实施例提供的图像特征提取装置100的组成示意图之一。在本申请实施例中，所述图像特征提取装置100可以包括特征图获取模块110、特征图预处理模块120、输入特征切割模块130、子图边界填充模块140、特征卷积运算模块150、子图边界叠加模块160及输出特征提取模块170。
特征图获取模块110，用于获取针对特征提取卷积神经网络的目标卷积核的待处理特征图。
特征图预处理模块120，用于根据预设图像尺寸对所述待处理特征图进行图像预处理，得到对应的输入特征图，其中所述输入特征图的图像尺寸为所述预设图像尺寸的整数倍；
输入特征切割模块130，用于按照所述预设图像尺寸将所述输入特征图切割为多个输入特征子图，其中每个所述输入特征子图的图像尺寸与所述预设图像尺寸保持一致。
子图边界填充模块140，用于针对每个所述输入特征子图，根据所述目标卷积核的卷积核尺寸对该输入特征子图进行边界填充，得到对应的待卷积特征子图。
特征卷积运算模块150，用于针对每个所述待卷积特征子图，基于Karatsuba算法采用所述目标卷积核对该待卷积特征子图进行卷积运算，得到对应的目标特征子图。
子图边界叠加模块160，用于针对得到的多个所述目标特征子图，根据所述输入特征图中各个所述输入特征子图的分布状况，将对应位置相邻的所述目标特征子图进行边界叠加，得到重叠特征图。
输出特征提取模块170，用于根据所述目标卷积核的卷积核尺寸对所述重叠特征图进行边界环切，得到对应的输出特征图，其中所述输出特征图用于表示与所述待处理特征图匹配的所述输入特征图的特征提取结果。
可选地，请参照图14，图14是本申请实施例提供的图像特征提取装置100的组成示意图之二。在本申请实施例中，所述图像特征提取装置100还可以包括切割尺寸配置模块180。
切割尺寸配置模块180，用于根据所述目标卷积核的卷积核尺寸对所述预设图像尺寸进行配置，其中所述预设图像尺寸所对应的行数大于或等于所述卷积核尺寸所对应的行数，所述预设图像尺寸所对应的列数大于或等于所述卷积核尺寸所对应的列数。
需要说明的是，本申请实施例所提供的图像特征提取装置100，其基本原理及产生的技术效果与前述的图像特征提取方法相同。为简要描述，本实施例部分未提及之处，可参考上述的针对图像特征提取方法的描述内容。
可以理解的是，文中上述的元素对于图像来说，可以是该图像的图像像素，也可以是该图像在某个颜色通道(例如，R(Red，红色)通道、G(Green，绿色)通道或B(Blue，蓝色)通道)处的图像灰度，具体的元素表达情况可根据图像特征提取需求进行适应性调整。
在本申请所提供的实施例中，应该理解到，所揭露的装置和方法，也可以通过其它的方式实现。以 上所描述的装置实施例仅仅是示意性的，例如，附图中的流程图和框图显示了根据本申请的实施例的装置、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段或代码的一部分，所述模块、程序段或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现方式中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个连续的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或动作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
另外，在本申请各个实施例中的各功能模块可以集成在一起形成一个独立的部分，也可以是各个模块单独存在，也可以两个或两个以上模块集成形成一个独立的部分。所述功能如果以软件功能模块的形式实现并作为独立的产品销售或使用时，可以存储在一个可读存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本申请各个实施例所述方法的全部或部分步骤。而前述的可读存储介质包括：U盘、移动硬盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、磁碟或者光盘等各种可以存储程序代码的介质。
综上所述，在本申请提供的图像特征提取方法及装置、计算机设备和可读存储介质中，本申请在获取到特征提取卷积神经网络的目标卷积核的待处理特征图后，会按照预设图像尺寸将由该待处理特征图经图像预处理得到的输入特征图切割为多个输入特征子图，并对每个输入特征子图进行边界填充，得到对应的待卷积特征子图，而后针对每个待卷积特征子图，通过Karatsuba算法消减目标卷积核与该待卷积特征子图的卷积运算过程中的乘法器资源损耗，并有效抑制卷积运算过程中的加法器资源损耗的增长，得到对应的目标特征子图，接着根据输入特征图中各输入特征子图的分布状况，将对应位置相邻的目标特征子图进行边界叠加，得到重叠特征图，最后对该重叠特征图进行边界环切，得到对应的输出特征图，从而从整体上降低特征提取时的计算资源损耗，并提升特征提取效率。
以上所述，仅为本申请的各种实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应当以权利要求的保护范围为准。

Claims (10)

  1. 一种图像特征提取方法，其特征在于，所述方法包括：
    获取针对特征提取卷积神经网络的目标卷积核的待处理特征图；
    根据预设图像尺寸对所述待处理特征图进行图像预处理，得到对应的输入特征图，其中所述输入特征图的图像尺寸为所述预设图像尺寸的整数倍；
    按照所述预设图像尺寸将所述输入特征图切割为多个输入特征子图，其中每个所述输入特征子图的图像尺寸与所述预设图像尺寸保持一致；
    针对每个所述输入特征子图，根据所述目标卷积核的卷积核尺寸对该输入特征子图进行边界填充，得到对应的待卷积特征子图；
    针对每个所述待卷积特征子图，基于Karatsuba算法采用所述目标卷积核对该待卷积特征子图进行卷积运算，得到对应的目标特征子图；
    针对得到的多个所述目标特征子图，根据所述输入特征图中各个所述输入特征子图的分布状况，将对应位置相邻的所述目标特征子图进行边界叠加，得到重叠特征图；
    根据所述目标卷积核的卷积核尺寸对所述重叠特征图进行边界环切，得到对应的输出特征图，其中所述输出特征图用于表示与所述待处理特征图匹配的所述输入特征图的特征提取结果。
  2. 根据权利要求1所述的方法，其特征在于，所述针对每个所述输入特征子图，根据所述目标卷积核的卷积核尺寸对该输入特征子图进行边界填充，得到对应的待卷积特征子图的步骤，包括：
    根据所述目标卷积核的卷积核尺寸，计算单个输入特征子图所需空白元素环的第一目标数目，其中所述第一目标数目由所述卷积核尺寸所对应的行/列数减一得到；
    针对每个所述输入特征子图，以该输入特征子图为中心在该输入特征子图周围外扩地依次填充第一目标数目个空白元素环，得到对应的待卷积特征子图。
  3. 根据权利要求1所述的方法，其特征在于，所述针对每个所述待卷积特征子图，基于Karatsuba算法采用所述目标卷积核对该待卷积特征子图进行卷积运算，得到对应的目标特征子图的步骤，包括：
    针对每个所述待卷积特征子图，根据对应的输入特征子图中不同分布位置的特征元素值，基于Karatsuba算法构建该输入特征子图的行列特征表达式；
    针对所述目标卷积核，根据该目标卷积核中不同分布位置的卷积元素值，基于Karatsuba算法构建该目标卷积核的行列特征表达式；
    针对每个所述待卷积特征子图，基于Karatsuba算法将对应输入特征子图的行列特征表达式与所述目标卷积核的行列特征表达式的乘法运算过程进行拆分换算，得到对应的拆分换算结果；
    针对每个所述待卷积特征子图，将对应的拆分换算结果基于与目标卷积核对应的预设卷积映射分布状况进行元素表达，得到对应的目标特征子图。
  4. 根据权利要求1所述的方法，其特征在于，所述根据所述输入特征图中各个所述输入特征子图的 分布状况，将对应位置相邻的所述目标特征子图进行边界叠加，得到重叠特征图的步骤，包括：
    根据所述目标卷积核的卷积核尺寸，确定单个目标特征子图的待叠加元素环的第二目标数目，其中所述第二目标数目由所述卷积核尺寸所对应的行/列数减一得到；
    针对每个目标特征子图，由外往内地选取该目标特征子图的第二目标数目个子图元素环作为待叠加元素环，得到该目标特征子图的待叠加区域；
    按照所述输入特征图中各个所述输入特征子图的分布状况，对各目标特征子图进行图像拼接，得到对应的图像拼接结果；
    在所述图像拼接结果中将位置相邻的待叠加区域各自的接触边缘区域进行元素叠加，得到所述重叠特征图。
  5. 根据权利要求1所述的方法，其特征在于，所述根据所述目标卷积核的卷积核尺寸对所述重叠特征图进行边界环切，得到对应的输出特征图的步骤，包括：
    根据所述目标卷积核的卷积核尺寸，确定所述重叠特征图的待切除元素环的第三目标数目，其中所述第三目标数目由所述卷积核尺寸所对应的行/列数减一得到；
    由外往内地选取该重叠特征图的第三目标数目个特征图元素环作为待切除元素环，得到该重叠特征图的待切除区域；
    对该重叠特征图中的待切除区域进行图像切除，得到与所述输入特征图对应的所述输出特征图。
  6. 根据权利要求1-5中任意一项所述的方法，其特征在于，所述方法还包括：
    根据所述目标卷积核的卷积核尺寸对所述预设图像尺寸进行配置，其中所述预设图像尺寸所对应的行数大于或等于所述卷积核尺寸所对应的行数，所述预设图像尺寸所对应的列数大于或等于所述卷积核尺寸所对应的列数。
  7. 一种图像特征提取装置，其特征在于，所述装置包括：
    特征图获取模块，用于获取针对特征提取卷积神经网络的目标卷积核的待处理特征图；
    特征图预处理模块，用于根据预设图像尺寸对所述待处理特征图进行图像预处理，得到对应的输入特征图，其中所述输入特征图的图像尺寸为所述预设图像尺寸的整数倍；
    输入特征切割模块，用于按照所述预设图像尺寸将所述输入特征图切割为多个输入特征子图，其中每个所述输入特征子图的图像尺寸与所述预设图像尺寸保持一致；
    子图边界填充模块，用于针对每个所述输入特征子图，根据所述目标卷积核的卷积核尺寸对该输入特征子图进行边界填充，得到对应的待卷积特征子图；
    特征卷积运算模块，用于针对每个所述待卷积特征子图，基于Karatsuba算法采用所述目标卷积核对该待卷积特征子图进行卷积运算，得到对应的目标特征子图；
    子图边界叠加模块，用于针对得到的多个所述目标特征子图，根据所述输入特征图中各个所述输入特征子图的分布状况，将对应位置相邻的所述目标特征子图进行边界叠加，得到重叠特征图；
    输出特征提取模块，用于根据所述目标卷积核的卷积核尺寸对所述重叠特征图进行边界环切，得到对应的输出特征图，其中所述输出特征图用于表示与所述待处理特征图匹配的所述输入特征图的特征提取结果。
  8. 根据权利要求7所述的装置，其特征在于，所述装置还包括：
    切割尺寸配置模块，用于根据所述目标卷积核的卷积核尺寸对所述预设图像尺寸进行配置，其中所述预设图像尺寸所对应的行数大于或等于所述卷积核尺寸所对应的行数，所述预设图像尺寸所对应的列数大于或等于所述卷积核尺寸所对应的列数。
  9. 一种计算机设备，其特征在于，包括处理器和存储器，所述存储器存储有能够被所述处理器执行的计算机程序，所述处理器可执行所述计算机程序，实现权利要求1-6中任意一项所述的图像特征提取方法。
  10. 一种可读存储介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器执行时，实现权利要求1-6中任意一项所述的图像特征提取方法。
PCT/CN2022/078874 2021-08-11 2022-03-02 图像特征提取方法及装置、计算机设备和可读存储介质 WO2023015871A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/396,324 US20240185571A1 (en) 2021-08-11 2023-12-26 Image feature extraction method and apparatus, and computer device and readable storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110917990.5 2021-08-11
CN202110917990.5A CN113642589B (zh) 2021-08-11 2021-08-11 图像特征提取方法及装置、计算机设备和可读存储介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/396,324 Continuation-In-Part US20240185571A1 (en) 2021-08-11 2023-12-26 Image feature extraction method and apparatus, and computer device and readable storage medium

Publications (1)

Publication Number Publication Date
WO2023015871A1 true WO2023015871A1 (zh) 2023-02-16

Family

ID=78420739

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/078874 WO2023015871A1 (zh) 2021-08-11 2022-03-02 图像特征提取方法及装置、计算机设备和可读存储介质

Country Status (3)

Country Link
US (1) US20240185571A1 (zh)
CN (1) CN113642589B (zh)
WO (1) WO2023015871A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116364223A (zh) * 2023-05-26 2023-06-30 平安科技（深圳）有限公司 特征处理方法、装置、计算机设备及存储介质

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113642589B (zh) * 2021-08-11 2023-06-06 南方科技大学 图像特征提取方法及装置、计算机设备和可读存储介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110136056A (zh) * 2018-02-08 2019-08-16 华为技术有限公司 图像超分辨率重建的方法和装置
CN110555847A (zh) * 2019-07-31 2019-12-10 瀚博半导体(上海)有限公司 一种基于卷积神经网络的图像处理方法及装置
US20200151541A1 (en) * 2018-11-08 2020-05-14 Arm Limited Efficient Convolutional Neural Networks
US20200175313A1 (en) * 2018-12-03 2020-06-04 Samsung Electronics Co., Ltd. Method and apparatus with dilated convolution
CN111738276A (zh) * 2020-06-19 2020-10-02 浪潮(北京)电子信息产业有限公司 基于多核卷积神经网络的图像处理方法、装置及设备
CN113642589A (zh) * 2021-08-11 2021-11-12 南方科技大学 图像特征提取方法及装置、计算机设备和可读存储介质

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10891537B2 (en) * 2019-03-20 2021-01-12 Huawei Technologies Co., Ltd. Convolutional neural network-based image processing method and image processing apparatus
CN110188795B (zh) * 2019-04-24 2023-05-09 华为技术有限公司 图像分类方法、数据处理方法和装置
CN110473137B (zh) * 2019-04-24 2021-09-14 华为技术有限公司 图像处理方法和装置
CN111260037B (zh) * 2020-02-11 2023-10-13 深圳云天励飞技术股份有限公司 图像数据的卷积运算方法、装置、电子设备及存储介质

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110136056A (zh) * 2018-02-08 2019-08-16 华为技术有限公司 图像超分辨率重建的方法和装置
US20200151541A1 (en) * 2018-11-08 2020-05-14 Arm Limited Efficient Convolutional Neural Networks
US20200175313A1 (en) * 2018-12-03 2020-06-04 Samsung Electronics Co., Ltd. Method and apparatus with dilated convolution
CN110555847A (zh) * 2019-07-31 2019-12-10 瀚博半导体(上海)有限公司 一种基于卷积神经网络的图像处理方法及装置
CN111738276A (zh) * 2020-06-19 2020-10-02 浪潮(北京)电子信息产业有限公司 基于多核卷积神经网络的图像处理方法、装置及设备
CN113642589A (zh) * 2021-08-11 2021-11-12 南方科技大学 图像特征提取方法及装置、计算机设备和可读存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116364223A (zh) * 2023-05-26 2023-06-30 平安科技（深圳）有限公司 特征处理方法、装置、计算机设备及存储介质
CN116364223B (zh) * 2023-05-26 2023-08-29 平安科技（深圳）有限公司 特征处理方法、装置、计算机设备及存储介质

Also Published As

Publication number Publication date
CN113642589B (zh) 2023-06-06
CN113642589A (zh) 2021-11-12
US20240185571A1 (en) 2024-06-06

Similar Documents

Publication Publication Date Title
WO2023015871A1 (zh) 图像特征提取方法及装置、计算机设备和可读存储介质
JP6431245B1 (ja) エッジ認識双方向画像処理
WO2021109876A1 (zh) 图像处理方法、装置、设备及存储介质
CN113034358B (zh) 一种超分辨率图像处理方法以及相关装置
WO2015143624A1 (en) Methods and systems for denoising images
CN112218161B (zh) 一种视频加密方法、装置及设备
TW202331599A (zh) 用於分層權重稀疏卷積處理之方法和系統及其相關的非暫時性電腦可讀儲存媒體
JP2023541350A (ja) 表畳み込みおよびアクセラレーション
CN114004754B (zh) 一种基于深度学习的场景深度补全系统及方法
JP7282474B2 (ja) 暗号化マスク確定方法、暗号化マスク確定装置、電子機器、記憶媒体およびコンピュータプログラム
CN112270332A (zh) 一种基于子流稀疏卷积的三维目标检测方法及系统
CN112465122B (zh) 优化神经网络模型中原始维度算子的装置和方法
WO2019209751A1 (en) Superpixel merging
Dinesh et al. Super-resolution of 3D color point clouds via fast graph total variation
CN109272526B (zh) 图像处理方法、系统及电子设备
CN111161289B (zh) 图像中物体轮廓精度提升方法、装置
CN107862316B (zh) 卷积运算方法及装置
CN113947521A (zh) 基于深度神经网络的图像分辨率转换方法及装置、终端设备
CN109685902B (zh) 建筑物模型效果生成方法及装置
CN105389580B (zh) 基于常量时间复杂度的图像矩形邻域极大或极小值计算方法
JP6361195B2 (ja) 画像処理装置、画像処理方法、画像処理プログラム及び記録媒体
CN111259165B (zh) 一种多核心元素的知识图谱边图布局方法及系统
Lin et al. Image inpainting using multiscale salient structure propagation
CN113469911B (zh) 图像去噪方法及装置、计算机可读存储介质、终端
WO2023185317A1 (zh) 虚拟地形的光照渲染方法、装置、介质、设备和程序产品

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22854876

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE