CN110826694B - Image processing method and device based on convolutional neural network - Google Patents
Image processing method and device based on convolutional neural network Download PDFInfo
- Publication number
- CN110826694B CN110826694B CN201911041729.2A CN201911041729A CN110826694B CN 110826694 B CN110826694 B CN 110826694B CN 201911041729 A CN201911041729 A CN 201911041729A CN 110826694 B CN110826694 B CN 110826694B
- Authority
- CN
- China
- Prior art keywords
- convolution
- screening
- value
- layer operation
- information
- Prior art date
- Legal status (The legal status 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 status listed.)
- Active
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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The application provides an image processing method and device based on a convolutional neural network, the technology can execute a first convolutional layer operation according to preset convolutional operation strategy information, and judge whether second convolutional layer operation information to be executed next meets the conditions: the convolution operation size is 1 x 1 and the step length value s is larger than 1; under the condition that the condition is met, screening out an effective value part corresponding to the second convolutional layer operation from the first convolutional layer operation result before the second convolutional layer operation is executed according to the step length value; and executing the convolution operation of the second convolution layer 1 x 1 on the effective value part according to the step length value of 1. Therefore, in the process of executing convolution operation to process the image, the corresponding effective value is screened out before the convolution with the convolution operation size of 1 x 1 and the step length value s larger than 1 is executed, the operation on an invalid value part is avoided, the image processing efficiency is improved, and the overall performance of the image processing chip is enhanced.
Description
Technical Field
The present application relates to the field of image processing, and in particular, to a technique for image processing based on a convolutional neural network.
Background
Convolutional Neural Network (CNN), which is one of artificial Neural networks, has been a research hotspot in the field of current image processing. The convolution neural network is a multilayer neural network, convolution operation is carried out on three-dimensional tensor data and a plurality of convolution kernels in each layer to extract image characteristics, the output of the image characteristics generated by the three-dimensional tensor data and each convolution kernel is two-dimensional tensor data, and a plurality of two-dimensional tensors output by the layer are organized together to form three-dimensional tensor data input of the next layer of neural network. Since the neural network may have a plurality of layers, the convolution operation amount of the convolutional neural network is very large.
At present, with the increase of the complexity of the convolution operation of image processing, the convolution operation with the convolution kernel size of 1 × 1 and the step length value greater than 1 occurs in the convolution operation process of image processing, a large amount of data is generated when the convolution layer operation is executed on the image to be processed in each neural network layer, and currently, the data in all the convolution operation processes are not effectively distinguished and screened according to needs, so that the bandwidth requirement on an internal/external memory of a processor is very high, and the overall performance of a convolution neural network system is influenced.
In view of the above problems, no effective solution has been proposed.
Disclosure of Invention
An object of the present application is to provide an image processing method and apparatus based on a convolutional neural network for convolutional operation with a convolutional kernel size of 1 x 1 and a step length value greater than 1, so as to solve the problems of distinguishing and screening validity of convolutional layer operation data results in the neural network as required, reducing bandwidth of a convolutional neural network processor memory, optimizing performance of the convolutional neural network processor, and the like in an image processing process.
According to an aspect of the present application, there is provided an image processing method based on a convolutional neural network, the method including:
executing the first convolution layer operation according to the preset convolution operation strategy information, and judging whether the second convolution layer operation information to be executed next meets the following conditions:
the convolution operation size is 1 x 1 and the step length value s is larger than 1;
under the condition that the condition is met, before the second convolutional layer operation is executed, screening out an effective value part corresponding to the second convolutional layer operation in the first convolutional layer operation result according to preset first effective value position information and the stepping length value corresponding to the second convolutional layer operation;
and executing the convolution operation of the second convolution layer 1 x 1 on the effective value part according to the step length value of 1.
According to another aspect of the present application, there is provided an image processing apparatus based on a convolutional neural network, the apparatus including:
a convolution operation information detection unit, configured to execute a first convolution layer operation according to preset convolution operation policy information, and determine whether second convolution layer operation information to be executed next satisfies the following condition:
the convolution operation size is 1 x 1 and the step length value s is larger than 1;
a first convolution operation result screening unit, configured to, when the condition is satisfied, screen, according to preset first effective value position information and the step length value corresponding to the second convolution layer operation, an effective value portion corresponding to the second convolution layer operation in the first convolution layer operation result before the second convolution layer operation is executed;
and a second convolution operation execution unit, configured to execute a second convolution layer 1 × 1 convolution operation on the effective value portion according to a step length value of 1.
Compared with the prior art, the method and the device can execute the first convolutional layer operation according to the preset convolutional operation strategy information, and judge whether the second convolutional layer operation information to be executed next meets the conditions: the convolution operation size is 1 x 1 and the step length value s is larger than 1; under the condition that the condition is met, screening out an effective value part corresponding to the second convolutional layer operation from the first convolutional layer operation result before the second convolutional layer operation is executed according to the step length value; and executing the convolution operation of the second convolution layer 1 x 1 on the effective value part according to the step length value of 1. Therefore, in the process of executing convolution operation to process the image, the corresponding effective value is screened out before the convolution with the convolution operation size of 1 x 1 and the step length value s larger than 1 is executed, the operation on an invalid value part is avoided, the image processing efficiency is improved, the overall performance of an image processing chip is enhanced, the Load/Store operation on invalid data is reduced, and the precious memory bandwidth is saved.
Drawings
Other features, objects and advantages of the present application will become more apparent upon reading of the following detailed description of non-limiting embodiments thereof, made with reference to the accompanying drawings in which:
FIG. 1 is a flowchart of a convolutional neural network-based image processing method according to an embodiment of the present application;
FIG. 2 is a diagram illustrating a screening of significant values in a result of a first convolution operation according to an embodiment of the present application;
FIG. 3 is a diagram illustrating a screening of significant values in a result of a first convolution operation according to another embodiment of the present application;
FIG. 4 is a diagram illustrating a screening of significant values in a result of a first convolution operation according to another embodiment of the present application;
fig. 5 shows a schematic block diagram of an image processing apparatus based on a convolutional neural network according to an embodiment of the present application.
The same or similar reference numbers in the drawings identify the same or similar elements.
Detailed Description
In order to make those skilled in the art better understand the technical solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only partial embodiments of the present application, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
It should be noted that the terms "first," "second," and the like in the description and claims of this application and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the application described herein are capable of operation in sequences other than those illustrated or described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.
The application provides an image processing technology based on a convolutional neural network, wherein the image processing comprises but is not limited to image processing modes such as image recognition, image segmentation, image classification, image optimization, image enhancement and the like. Specifically, in the image processing process, a convolutional neural network for image processing is set for the factors such as the image to be processed and the recognition purpose, that is, a convolutional operation strategy is executed for the image to be processed, and the image is recognized, segmented, classified, optimized, enhanced and the like by executing the corresponding convolutional neural network. With the development of convolutional neural network technology, operations using convolution kernels with smaller sizes are more and more, and especially, convolution kernel operations with the size of 1 × 1 are applied in a large amount in actual image processing. This application is mainly directed at when carrying out the operation that the convolution layer size is 1 × 1 step length and is greater than 1, also when carrying out this type of convolution operation, in the output result of its preceding convolution layer, not all values are all effectual, in order to avoid carrying out this convolution layer operation to the invalid value, thereby increase image processing chip's processing burden, this application just screens the effective value that corresponds before carrying out this convolution operation, thereby effectively improve image processing convolution neural network's treatment effeciency, and the wholeness ability of image processing chip has been improved.
According to one aspect of the present application, there is provided an image processing method based on a convolutional neural network. Referring to fig. 1, the method includes the steps of:
s1 executes the first convolutional layer operation according to the preset convolutional operation policy information, and determines whether the second convolutional layer operation information to be executed next satisfies the following condition:
the convolution size is 1 x 1 and the step length value s is greater than 1.
In this step, the convolution operation policy information is convolution neural network information executed on the image to be processed, and includes, for example, but not limited to, information such as the number of layers of convolution operation, the size of each layer of convolution operation, and the step length value thereof, which is preset before the image to be processed. The first convolutional layer is a convolutional layer in a current execution process or a preset convolutional neural network after execution is finished, and the second convolutional layer is a convolutional layer which is executed continuously correspondingly after the first convolutional layer is finished; the length of convolution step is a natural number, which may be 1, 2, or 4.
Specifically, in this embodiment, the number of convolution layers included in the entire convolutional neural network, the convolution size of each convolutional layer, the step length, and the like are set in advance, and compilation of instructions is performed by a compiler. The compiler judges whether the second convolutional layer after the current convolutional layer operation is executed satisfies the following conditions according to preset convolutional operation strategy information: the convolution size is 1 x 1 and the step length value s is greater than 1. If the second convolution layer operation information meets the condition; and if the result indicates that an invalid value which is not needed by the operation of the second convolutional layer exists in the operation result of the currently executed convolutional layer, the compiler should start a corresponding instruction, screen the valid value in the current convolutional operation result according to preset convolutional strategy information, and input the valid value part to the second convolutional layer only, so that the efficiency of the convolutional operation is improved.
Here, the information such as the size of the convolutional layer and the step length in the second convolutional layer operation information is preset information, and the determination method is not limited at all, for example, the comparison may be performed with the standard convolutional size and step length value information, and the determination is performed according to the comparison result, or the convolutional layer information meeting the above condition is preset in advance when the compiler compiles the whole convolutional layer operation neural network, and the compiler instructs the convolutional neural network to execute the above convolutional layer meeting the above condition in the operation process, and then starts the corresponding module to perform the determination.
S2, if the above condition is satisfied, before executing the second convolutional layer operation, screens out an effective value part corresponding to the second convolutional layer operation in the first convolutional layer operation result according to the preset first effective value position information and the step length value corresponding to the second convolutional layer operation.
Specifically, when the size of the second convolution layer is 1 × 1 and the step length value S is greater than 1, all contents in the operation result output after the first convolution layer operation is completed are not used as valid input information when the second convolution layer operation is executed, and even if all the operation results of the first convolution layer are directly input to the second convolution layer as input, the second convolution layer operation performs convolution operation on all input data (including valid and invalid portions), but subsequently, invalid portions in the second convolution operation result still need to be discarded.
In this step, the first effective value is the effective value arranged at the top left corner of the first convolutional layer operation result, as shown in fig. 2 to 4, where the "√" sign in fig. 2-1, 3-1, and 4-1 indicates the position of the first effective value. And under the condition that the second convolution layer operation information meets the requirement that the size is 1 x 1 and the step length value is greater than 1, screening effective values of a first convolution layer operation result to be input as a second convolution layer operation before executing the second convolution layer operation, thereby improving the image processing efficiency and saving operation resources. Further, according to the position information of the first effective value of the second convolution layer in the first convolution layer operation result in the preset convolution operation strategy information, the initial position of the screening can be determined, and then according to the step length information of the second convolution layer operation, the screening mode is determined, for example, if the step length value is 2, then from the initial position, one effective value is screened every other 1 position according to the preset sequence, and if the step length is 4, then from the initial position, one effective value is screened every other 3 positions according to the preset sequence. Here, when the position information of the first effective value and the step length value corresponding to the preset second convolution layer operation are known, the method of screening out the effective value is not limited, and may be implemented by a screening circuit hardware, a screening software module, or the like; the specific implementation manner of the hardware of the screening circuit is not limited, and for example, the hardware may be a filtering circuit; the specific implementation manner of the software module is not limited, and the software module which can be matched with the compiler and can realize the effective value screening in the first convolution operation result now and in the future is within the protection scope of the present application. Accordingly, the specific manner of initiating the screening is not limited, and for example, the specific manner may be a prompt message of a compiler or an issue of a related instruction, or a call of a corresponding software screening module. Specifically, the work efficiency of the screening mode realized by matching screening circuit hardware with the compiler is higher, the specific implementation mode of realizing screening by calling related software modules through the compiler may be more flexible, but the processing efficiency may be lower than that of screening circuit hardware.
S3 performs the second convolution layer 1 x 1 convolution operation on the significant value portion according to a step length value of 1.
In this step, the step length value of the second convolutional layer operation is considered when the effective value is screened, where the effective value is that the convolution operation of the second convolutional layer for the effective value is directly executed according to a preset convolution size 1 × 1 to complete the second convolutional layer operation.
In some embodiments, prior to performing the second convolutional layer operation comprises one of:
(1) after executing the first convolution layer operation, outputting the corresponding convolution operation result to the second convolution layer; in this case, for the first convolutional layer operation being completed but not yet output to the second convolutional layer, the filtering of the effective value may be inside the first convolutional layer or outside the first convolutional layer, and is not limited specifically.
(2) After the output result of the first convolution layer is obtained, before the second convolution layer operation is executed; at this time, the operation of the first convolutional layer is completed and the operation result is output to the second convolutional layer, and before the operation of the second convolutional layer is performed, the screening of the effective value should be completed in the second convolutional layer at this time.
In some embodiments, a valid value screening module screens out a valid value portion of the first convolutional layer operation result corresponding to the execution of the second convolutional layer operation.
Specifically, the effective value screening module needs to screen the effective value according to the position of the first effective value of the preset second convolution layer operation in the first convolution layer operation result and the size of the step value S of the second convolution layer, the specific implementation manner is not limited, and the effective value screening module can be implemented in a manner of software/hardware or combination of the software and the hardware, specifically can be implemented by matching corresponding instruction information of a screening circuit hardware with a compiler, and can also be implemented by using a screening signal called by the compiler through a software module.
In some embodiments, the S2 includes:
s21 (not shown) determines, according to the preset position information of the first effective value of the second convolution layer in the first convolution layer operation result, a starting position where the effective value screening module starts screening, that is, a position of the first effective value in the first convolution layer operation result. Specifically, referring to fig. 2, taking effective value screening by screening circuit hardware as an example, the screening module performs effective value screening according to a preset sequence from a starting position according to the step length value S, where S is 2 in fig. 2-1, a first effective value position is a first row and a first column position, and fig. 2-2 is an effective value screened by the screening module.
In this step, in order to ensure that all valid values required for the second convolutional layer operation are screened from the first convolutional layer operation result, it is necessary to consider a case where the first convolutional layer operation result is extended, specifically, as shown in fig. 3 and 4. Specifically, whether the first convolution layer operation result needs to be extended or not and size information that needs to be extended need to be determined according to position information of the first effective value in the first convolution layer operation result, step length information of the second convolution layer operation, and the like.
S22 (not shown), the valid value filtering module performs non-overlapping filtering on the first convolution layer operation result from the start position; specifically, referring to fig. 2, the screening process of the effective value screening module may be implemented by a screening circuit hardware or a software module, and the screening is performed from left to right and from top to bottom according to a preset sequence, where the screening is non-overlapping and traversing screening, so as to ensure that all effective values are screened, and no repeated screening wastes computation resources.
S23 (not shown) determines a valid value portion of the first convolutional layer operation result corresponding to the second convolutional layer operation according to the filtering result. Specifically, as shown in 2-3 of FIG. 2, after the valid values are selected, the valid values are gathered together again according to their original positions and used as the input of the second convolutional layer operation.
In some embodiments, prior to the S22, the S2 further comprises:
s24 (not shown) determines whether the size of the first convolution layer operation result needs to be extended according to the size value of the first convolution layer operation result, the start position information, and the effective value screening size.
In order to ensure that all the significant values required for the second convolutional layer operation are screened from the first convolutional layer operation result, taking the screening of the significant values by the screening circuit hardware as an example, it is necessary to consider the case of extending the first convolutional layer operation result, specifically, as shown in fig. 3 and 4. Specifically, the actual area information of the first convolution layer operation result for screening is determined according to the first convolution layer operation result and the first effective value position information (i.e. the initial position information), whether the first convolution layer operation result needs to be extended is determined according to the actual area information and the step length information of the second convolution layer operation, as shown in fig. 3, the step value of the second convolution layer operation in fig. 3-2 is 2, and the position of the first effective value in the first convolution layer operation result is indicated by "√" in fig. 3-1, so that the thickened line-bar sub-area is determined to be the actual screening area of the first convolution layer operation result; further, according to the principle that whether the deterministic input can be obtained by each operation of the screening circuit in the actual screening area through the step length value, whether the actual screening area needs to be extended is judged.
In some embodiments, in a case that the first convolution layer operation result size needs to be extended, the S2 further includes:
s25 (not shown) determines, according to the size value of the first convolution layer operation result and the start position information of the start filtering, the actual filtering area size information of the first convolution layer operation result by the valid value filtering module.
S26 (not shown) determines size information that needs to be extended for the first convolution layer operation result according to the size information of the actual filtering area and the filtering size of the effective value filtering module.
Specifically, the actual area information of the first convolution layer operation result for screening is determined according to the first convolution layer operation result and the first effective value position information (i.e. the initial position information), whether the first convolution layer operation result needs to be extended is determined according to the actual area information and the step length information of the second convolution layer operation, as shown in fig. 3, the step value of the second convolution layer operation in fig. 3-2 is 2, and the position of the first effective value in the first convolution layer operation result is indicated by "√" in fig. 3-1, so that the thickened line-bar sub-area is determined to be the actual screening area of the first convolution layer operation result; further, according to the principle that whether the deterministic input can be obtained by each operation of the screening circuit in the actual screening area through the step length value, whether the actual screening area needs to be extended is judged.
If the actual screening area of the first convolution operation result is screened according to the step length of the second convolution layer operation as the screening size, if the screening circuit in the actual screening area can obtain the deterministic input each time, no extension is needed, otherwise, extension is needed, as shown in figure 2, the step length of the second convolution layer operation is 2, the actual screening area is 4, all the actual screening areas can be traversed, extension is not needed, as shown in figure 2-2, as shown in figures 3 and 4, in order to screen the actual screening area of the first convolution operation result according to the step length of the second convolution layer operation as the screening size, when the screening circuit screens the actual screening area according to the step length value, if the input size (length, width) of the actual screening area does not accord with the integral multiple of the size (namely the step value) of the screening circuit, the actual screening area needs to be extended to ensure that all values within the actual screening area are screened, as shown in fig. 3-2 and 4-2.
In some embodiments, the S2 further includes:
s27 (not shown) determines extension area information of the operation result size of the first convolution layer according to the size information to be extended, where the extension area information includes column number information extending to the right side and/or row number information extending to the lower side.
In some embodiments, the column value extending to the right is the remainder of the division of the number of columns of the actual screening area size by the step length s.
And the downward extension line numerical value is the remainder of division of the line number of the actual screening area size and the stepping length s.
In some embodiments, the S22 includes:
and the effective value screening module carries out non-overlapping screening on the first convolution layer operation result from top to bottom from left to right in sequence from the initial position in the first convolution layer operation result.
In some embodiments, the value in the extension area is set to an invalid value, for example, may be set to 0, or may be other values preset to be invalid.
In some embodiments, the method further comprises:
s4 (not shown) presets convolution layer operation strategy information in the convolutional network, where the convolution layer operation strategy information at least includes convolution layer operation size information and step length value information corresponding to the convolution layer operation size information.
Specifically, the number of convolution layers, the convolution size of each convolution layer, the step length value, the first effective value, and the like included in the convolutional neural network need to be preset as needed before the corresponding image is processed.
Compared with the prior art, the method and the device can execute the first convolutional layer operation according to the preset convolutional operation strategy information, and judge whether the second convolutional layer operation information to be executed next meets the conditions: the convolution operation size is 1 x 1 and the step length value s is larger than 1; under the condition that the condition is met, screening out an effective value part corresponding to the second convolutional layer operation from the first convolutional layer operation result before the second convolutional layer operation is executed according to the step length value; and executing the convolution operation of the second convolution layer 1 x 1 on the effective value part according to the step length value of 1. Therefore, in the process of executing convolution operation to process the image, the corresponding effective value is screened out before the convolution with the convolution operation size of 1 x 1 and the step length value s larger than 1 is executed, the operation on an invalid value part is avoided, the image processing efficiency is improved, the overall performance of an image processing chip is enhanced, the Load/Store operation on invalid data is reduced, and the precious memory bandwidth is saved.
According to an aspect of the present application, there is provided an image processing apparatus 100 based on a convolutional neural network, as shown in fig. 5, the apparatus including:
a convolution operation information detection unit 110, configured to execute a first convolution layer operation according to preset convolution operation policy information, and determine whether second convolution layer operation information to be executed next satisfies the following condition:
the convolution size is 1 x 1 and the step length value s is greater than 1.
Specifically, the convolution operation policy information is convolution neural network information executed on the image to be processed, and includes, for example, but not limited to, information such as the number of layers of convolution operation, the size of each layer of convolution operation, and the step length value thereof, which is preset before the image to be processed. The first convolutional layer is a convolutional layer in a current execution process or a preset convolutional neural network after execution is finished, and the second convolutional layer is a convolutional layer which is executed continuously correspondingly after the first convolutional layer is finished; the length of convolution step is a natural number, which may be 1, 2, or 4.
Specifically, in this embodiment, the number of convolution layers included in the entire convolutional neural network, the convolution size of each convolutional layer, the step length, and the like are set in advance, and compilation of instructions is performed by a compiler. The compiler judges whether the second convolutional layer after the current convolutional layer operation is executed satisfies the following conditions according to preset convolutional operation strategy information: the convolution size is 1 x 1 and the step length value s is greater than 1. If the second convolution layer operation information meets the condition; and if the result indicates that an invalid value which is not needed by the operation of the second convolutional layer exists in the operation result of the currently executed convolutional layer, the compiler should start a corresponding instruction, screen the valid value in the current convolutional operation result according to preset convolutional strategy information, and input the valid value part to the second convolutional layer only, so that the efficiency of the convolutional operation is improved.
Here, the information such as the size of the convolutional layer and the step length in the second convolutional layer operation information is preset information, and the determination method is not limited at all, for example, the comparison may be performed with the standard convolutional size and step length value information, and the determination is performed according to the comparison result, or the convolutional layer information meeting the above condition is preset in advance when the compiler compiles the whole convolutional layer operation neural network, and the compiler instructs the convolutional neural network to execute the above convolutional layer meeting the above condition in the operation process, and then starts the corresponding module to perform the determination.
A first convolution operation result screening unit 120, configured to, if the above condition is met, screen out, according to preset first effective value position information and the step length value corresponding to the second convolution layer operation, an effective value portion corresponding to the second convolution layer operation in the first convolution layer operation result before the second convolution layer operation is executed.
Specifically, when the size of the second convolution layer is 1 × 1 and the step length value S is greater than 1, all contents in the operation result output after the first convolution layer operation is completed are not used as valid input information when the second convolution layer operation is executed, and even if all the operation results of the first convolution layer are directly input to the second convolution layer as input, the second convolution layer operation performs convolution operation on all input data (including valid and invalid portions), but subsequently, invalid portions in the second convolution operation result still need to be discarded.
Specifically, the first effective value is the effective value arranged at the top left corner of the first convolution layer operation result, as shown in fig. 2 to 4, where the "√" numeral in fig. 2-1, 3-1, and 4-1 indicates the position of the first effective value. And under the condition that the second convolution layer operation information meets the requirement that the size is 1 x 1 and the step length value is greater than 1, screening effective values of a first convolution layer operation result to be input as a second convolution layer operation before executing the second convolution layer operation, thereby improving the image processing efficiency and saving operation resources. Further, according to the position information of the first effective value of the second convolution layer in the first convolution layer operation result in the preset convolution operation strategy information, the initial position of the screening can be determined, and then according to the step length information of the second convolution layer operation, the screening mode is determined, for example, if the step length value is 2, then from the initial position, one effective value is screened every other 1 position according to the preset sequence, and if the step length is 4, then from the initial position, one effective value is screened every other 3 positions according to the preset sequence. Here, when the position information of the first significant value and the step length value corresponding to the preset second convolution layer operation are known, the method of screening out the significant value is not limited, and may be implemented by a screening circuit hardware, a screening software module, or the like; the specific implementation manner of the hardware of the screening circuit is not limited, and for example, the hardware may be a filtering circuit; the specific implementation manner of the software module is not limited, and the software module which can be matched with the compiler and can realize the effective value screening in the first convolution operation result now and in the future is within the protection scope of the present application. Accordingly, the specific manner of initiating the screening is not limited, and for example, the specific manner may be a prompt message of a compiler or an issue of a related instruction, or a call of a corresponding software screening module.
Specifically, the work efficiency of the screening mode realized by matching screening circuit hardware with the compiler is higher, the specific implementation mode of realizing screening by calling related software modules through the compiler may be more flexible, but the processing efficiency may be lower than that of screening circuit hardware.
A second convolution operation execution unit 130, configured to execute the second convolution layer 1 × 1 convolution operation on the effective value portion according to a step length value of 1.
Specifically, the step length value of the second convolutional layer operation is considered when the effective value is screened, where the effective value is that the second convolutional layer directly executes the convolutional operation on the effective value according to a preset convolutional size to complete the second convolutional layer operation.
In some embodiments, prior to performing the second convolutional layer operation comprises one of:
after executing the first convolution layer operation, outputting the corresponding convolution operation result to the second convolution layer; in this case, for the first convolutional layer operation being completed but not yet output to the second convolutional layer, the filtering of the effective value may be inside the first convolutional layer or outside the first convolutional layer, and is not limited specifically.
After the output result of the first convolution layer is obtained, before the second convolution layer operation is executed; at this time, the operation of the first convolutional layer is completed and the operation result is output to the second convolutional layer, and before the operation of the second convolutional layer is performed, the screening of the effective value should be completed in the second convolutional layer at this time.
In some embodiments, a valid value screening module screens out a valid value portion of the first convolutional layer operation result corresponding to the execution of the second convolutional layer operation.
Specifically, the effective value screening module needs to screen the effective value according to the position of the first effective value of the preset second convolution layer operation in the first convolution layer operation result and the size of the step value S of the second convolution layer, and the specific implementation manner is not limited and may be implemented by software/hardware or a combination of the software and the hardware, for example, the effective value screening module may be implemented by matching the screening circuit hardware with corresponding instruction information of the compiler, or may be implemented by using a start screening signal called by the compiler through the software module, and the like.
In some embodiments, the first convolution operation result filtering unit 120 includes:
a screening start position determining module (not shown) configured to determine, according to a position of a preset first effective value of the second convolution layer in the first convolution layer operation result, a start position at which the effective value screening module starts screening, where a screening size value of the effective value screening module is the step length s value.
Specifically, referring to fig. 2, taking the screening of the valid values by the screening circuit hardware as an example, the screening module performs the valid value screening according to the step length value S in a preset sequence from the start position, where S is 2 in fig. 2-1, the first valid value position is the first row and the first column position, and fig. 2-2 is the valid value screened by the screening module.
Specifically, in order to ensure that all valid values required for the second convolutional layer operation are screened from the first convolutional layer operation result, it is necessary to consider a case where the first convolutional layer operation result is extended, specifically, as shown in fig. 3 and 4. Specifically, whether the first convolution layer operation result needs to be extended or not and size information that needs to be extended need to be determined according to position information of the first effective value in the first convolution layer operation result, step length information of the second convolution layer operation, and the like.
A valid value screening module (not shown) for performing non-overlapping screening on the first convolution layer operation result from the start position. Specifically, referring to fig. 2, the screening process of the effective value screening module may be implemented by a screening circuit hardware or a software module, and the screening is performed from left to right and from top to bottom according to a preset sequence, where the screening is non-overlapping and traversing screening, so as to ensure that all effective values are screened, and no repeated screening wastes computation resources.
And an effective value determining module (not shown) configured to determine, according to the filtering result of the effective value filtering module, an effective value portion of the first convolutional layer operation result that is required by the second convolutional layer operation. Specifically, as shown in 2-3 of FIG. 2, after the valid values are selected, the valid values are gathered together again according to their original positions and used as the input of the second convolutional layer operation.
In some embodiments, the first convolution operation result filtering unit 120 further includes:
and the extension judging module is used for judging whether the operation result size of the first convolution layer needs to be extended or not according to the position information of the first effective value of the second convolution layer and the step length value of the second effective value.
In order to ensure that all the significant values required for the second convolutional layer operation are screened from the first convolutional layer operation result, taking the screening of the significant values by the screening circuit hardware as an example, it is necessary to consider the case of extending the first convolutional layer operation result, specifically, as shown in fig. 3 and 4. Specifically, the actual area information of the first convolution layer operation result for screening is determined according to the first convolution layer operation result and the first effective value position information (i.e. the initial position information), whether the first convolution layer operation result needs to be extended is determined according to the actual area information and the step length information of the second convolution layer operation, as shown in fig. 3, the step value of the second convolution layer operation in fig. 3-2 is 2, and the position of the first effective value in the first convolution layer operation result is indicated by "√" in fig. 3-1, so that the thickened line-bar sub-area is determined to be the actual screening area of the first convolution layer operation result; further, according to the principle that whether the deterministic input can be obtained by each operation of the screening circuit in the actual screening area through the step length value, whether the actual screening area needs to be extended is judged.
In some embodiments, in a case that the first convolution layer operation result size needs to be extended, the first convolution layer operation result screening unit further includes:
an actual screening area size determining module, configured to determine, according to the first effective value position information of the second convolution layer and the step length value thereof, actual screening area size information of the first convolution operation result by the effective value screening module;
and the extension size information determining module is used for determining size information which needs to be extended according to the size information of the actual screening area and the screening size of the screening module.
Specifically, the actual area information of the first convolution layer operation result for screening is determined according to the first convolution layer operation result and the first effective value position information (i.e. the initial position information), whether the first convolution layer operation result needs to be extended is determined according to the actual area information and the step length information of the second convolution layer operation, as shown in fig. 3, the step value of the second convolution layer operation in fig. 3-2 is 2, and the position of the first effective value in the first convolution layer operation result is indicated by "√" in fig. 3-1, so that the thickened line-bar sub-area is determined to be the actual screening area of the first convolution layer operation result; further, according to the principle that whether the deterministic input can be obtained by each operation of the screening circuit in the actual screening area through the step length value, whether the actual screening area needs to be extended is judged.
If the actual screening area of the first convolution operation result is screened according to the step length of the second convolution layer operation as the screening size, if the screening circuit in the actual screening area can obtain the deterministic input each time, no extension is needed, otherwise, extension is needed, as shown in figure 2, the step length of the second convolution layer operation is 2, the actual screening area is 4, all the actual screening areas can be traversed, extension is not needed, as shown in figure 2-2, as shown in figures 3 and 4, in order to screen the actual screening area of the first convolution operation result according to the step length of the second convolution layer operation as the screening size, when the screening circuit screens the actual screening area according to the step length value, if the input size (length, width) of the actual screening area does not accord with the integral multiple of the size (namely the step value) of the screening circuit, the actual screening area needs to be extended to ensure that all values within the actual screening area are screened, as shown in fig. 3-2 and 4-2.
In some embodiments, the first convolution operation result filtering unit further includes:
and the extension area information determining module is used for determining extension area information of the operation result size of the first convolution layer according to the size information needing to be extended, wherein the extension area information comprises column number information extending towards the right side and/or row number information extending towards the lower side.
In some embodiments, the column value extending to the right is a remainder of division of the number of columns of the actual screening area size by the step length s;
and the downward extension line numerical value is the remainder of division of the line number of the actual screening area size and the stepping length s.
In some embodiments, the first convolution operation result screening unit is configured to:
the effective value screening module carries out non-overlapping screening on the first convolution layer operation result from top to bottom from left to right in sequence from the initial position in the first convolution layer operation result.
In some embodiments, the value in the extension area is set to an invalid value, for example, may be set to 0, or may be other values preset to be invalid.
In some embodiments, the apparatus further comprises:
and the convolutional layer operation strategy information unit is used for presetting convolutional layer operation strategy information in the convolutional network, wherein the convolutional layer operation strategy information at least comprises convolutional layer operation size information and corresponding step length value information.
Specifically, the number of convolution layers, the convolution size of each convolution layer, the step length value, the first effective value, and the like included in the convolutional neural network need to be preset as needed before the corresponding image is processed.
Compared with the prior art, the method and the device can execute the first convolutional layer operation according to the preset convolutional operation strategy information, and judge whether the second convolutional layer operation information to be executed next meets the conditions: the convolution operation size is 1 x 1 and the step length value s is larger than 1; under the condition that the condition is met, screening out an effective value part corresponding to the second convolutional layer operation from the first convolutional layer operation result before the second convolutional layer operation is executed according to the step length value; and executing the convolution operation of the second convolution layer 1 x 1 on the effective value part according to the step length value of 1. Therefore, in the process of executing convolution operation to process the image, the corresponding effective value is screened out before the convolution with the convolution operation size of 1 x 1 and the step length value s larger than 1 is executed, the operation on an invalid value part is avoided, the image processing efficiency is improved, the overall performance of an image processing chip is enhanced, the Load/Store operation on invalid data is reduced, and the precious memory bandwidth is saved.
The above-mentioned serial numbers of the embodiments of the present application are merely for description and do not represent the merits of the embodiments.
In the above embodiments of the present application, the descriptions of the respective embodiments have respective emphasis, and for parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.
In the embodiments provided in the present application, it should be understood that the disclosed technology can be implemented in other ways. The above-described embodiments of the apparatus are merely illustrative, and for example, the division of the units may be a logical division, and in actual implementation, there may be another division, for example, multiple units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, units or modules, and may be in an electrical or other form.
The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.
In addition, functional units in the embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, and can also be realized in a form of a software functional unit.
The integrated unit, if implemented in the form of a software functional unit and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present application may be substantially implemented or contributed to by the prior art, or all or part of the technical solution may be embodied in a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a removable hard disk, a magnetic or optical disk, and other various media capable of storing program codes.
The foregoing is merely a preferred embodiment of the present application and it should be noted that, as would be apparent to one of ordinary skill in the art, the present application is not limited to the details of the above-described exemplary embodiments and that the present application may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. Any reference sign in a claim should not be construed as limiting the claim concerned. Furthermore, it is obvious that the word "comprising" does not exclude other elements or steps, and the singular does not exclude the plural. A plurality of units or means recited in the apparatus claims may also be implemented by one unit or means in software or hardware. The terms first, second, etc. are used to denote names, but not any particular order.
Claims (20)
1. An image processing method based on a convolutional neural network, wherein the method comprises the following steps:
executing the first convolution layer operation according to the preset convolution operation strategy information, and judging whether the second convolution layer operation information to be executed next meets the following conditions:
the convolution operation size is 1 x 1 and the step length value s is larger than 1;
under the condition that the second convolution layer operation is satisfied, screening out an effective value part corresponding to the second convolution layer operation in the first convolution layer operation result according to preset first effective value position information and the step length value corresponding to the second convolution layer operation before the second convolution layer operation is executed; the method specifically comprises the following steps:
determining an initial position of the effective value screening module for starting screening according to preset position information of a first effective value of the second convolution layer in an operation result of the first convolution layer, wherein a screening size value of the effective value screening module is the step length s value;
the effective value screening module starts to carry out non-overlapping screening on the first convolution layer operation result from the initial position;
determining the effective value part of the first convolution layer operation result corresponding to the second convolution layer operation according to the screening result
And executing the convolution operation of the second convolution layer 1 x 1 on the effective value part according to the step length value of 1.
2. The method of claim 1, wherein prior to performing the second convolutional layer operation comprises one of:
after executing the first convolution layer operation, outputting the corresponding convolution operation result to the second convolution layer;
and after the output result of the first convolution layer is obtained, executing the second convolution layer before operation.
3. The method of claim 1 or 2, wherein a significant value portion of the first convolutional layer operation result corresponding to the execution of the second convolutional layer operation is filtered out by a significant value filtering module.
4. The method of claim 3, wherein, before the valid value filtering module performs non-overlapping filtering on the first convolutional layer operation result from the start position, if the above condition is satisfied, before performing the second convolutional layer operation, filtering out a valid value portion corresponding to performing the second convolutional layer operation in the first convolutional layer operation result according to preset first valid value position information and the step length value corresponding to the second convolutional layer operation, further comprises:
and judging whether the operation result size of the first convolution layer needs to be extended or not according to the position information of the first effective value of the second convolution layer and the step length value of the second convolution layer.
5. The method of claim 4, wherein, in a case that the size of the first convolutional layer operation result needs to be extended, and in a case that the above condition is satisfied, before executing the second convolutional layer operation, screening out a valid value part corresponding to the execution of the second convolutional layer operation in the first convolutional layer operation result according to preset first valid value position information and the step length value corresponding to the second convolutional layer operation further comprises:
determining the actual screening area size information of the first convolution operation result by the effective value screening module according to the first effective value position information and the stepping length value of the second convolution layer;
and determining size information which needs to be extended in the first convolution layer operation result according to the size information of the actual screening area and the screening size of the screening module.
6. The method of claim 5, wherein if the condition is satisfied, screening out a valid value portion of the first convolutional layer operation result corresponding to the second convolutional layer operation according to preset first valid value position information and the step length value corresponding to the second convolutional layer operation before the second convolutional layer operation is executed further comprises:
and determining extension area information of the operation result size of the first convolution layer according to the size information needing to be extended, wherein the extension area information comprises column number information extending towards the right side and/or row number information extending towards the lower side.
7. The method of claim 6, wherein the column value extending to the right is a remainder of a division of the number of columns of the actual screening area size by the step length s;
and the downward extension line numerical value is the remainder of division of the line number of the actual screening area size and the stepping length s.
8. The method of any of claims 4 to 7, wherein the valid value filtering module performing non-overlapping filtering on the first convolution layer operation result starting from the start position comprises:
the effective value screening module carries out non-overlapping screening on the first convolution layer operation result from top to bottom from left to right in sequence from the initial position in the first convolution layer operation result.
9. The method according to claim 6 or 7, wherein the value in the extension area is set to an invalid value.
10. The method of any of claims 1, 2, and 4-7, wherein the method further comprises:
the method comprises the steps of presetting convolutional layer operation strategy information in a convolutional network, wherein the convolutional layer operation strategy information at least comprises convolutional layer operation size information and step length value information corresponding to the convolutional layer operation size information.
11. An image processing apparatus based on a convolutional neural network, wherein the apparatus comprises:
a convolution operation information detection unit, configured to execute a first convolution layer operation according to preset convolution operation policy information, and determine whether second convolution layer operation information to be executed next satisfies the following condition:
the convolution operation size is 1 x 1 and the step length value s is larger than 1;
a first convolution operation result screening unit, configured to, when the condition is satisfied, screen, according to preset first effective value position information and the step length value corresponding to the second convolution layer operation, an effective value portion corresponding to the second convolution layer operation in the first convolution layer operation result before the second convolution layer operation is executed; wherein the first convolution operation result screening unit includes:
a screening starting position determining module, configured to determine, according to a position of a preset first effective value of the second convolution layer in an operation result of the first convolution layer, a starting position at which the effective value screening module starts screening, where a screening size value of the effective value screening module is the step length s value;
the effective value screening module is used for carrying out non-overlapping screening on the first convolution layer operation result from the initial position;
an effective value determining module, configured to determine, according to the screening result, an effective value portion required by the second convolutional layer operation in the first convolutional layer operation result;
and a second convolution operation execution unit, configured to execute a second convolution layer 1 × 1 convolution operation on the effective value portion according to a step length value of 1.
12. The apparatus of claim 11, wherein prior to performing the second convolutional layer operation comprises one of:
after executing the first convolution layer operation, outputting the corresponding convolution operation result to the second convolution layer;
and after the output result of the first convolution layer is obtained, executing the second convolution layer before operation.
13. The apparatus according to claim 11 or 12, wherein a significant value part of the first convolutional layer operation result corresponding to the execution of the second convolutional layer operation is filtered out by a significant value filtering module.
14. The apparatus of claim 13, wherein the first convolution operation result filtering unit further comprises:
and the extension judging module is used for judging whether the operation result size of the first convolution layer needs to be extended or not according to the position information of the first effective value of the second convolution layer and the step length value of the second effective value.
15. The apparatus of claim 14, wherein in case the first convolution layer operation result size needs to be extended, the first convolution layer operation result screening unit further comprises:
an actual screening area size determining module, configured to determine, according to the first effective value position information of the second convolution layer and the step length value thereof, actual screening area size information of the first convolution operation result by the effective value screening module;
and the extension size information determining module is used for determining size information which needs to be extended according to the size information of the actual screening area and the screening size of the screening module.
16. The apparatus of claim 15, wherein the first convolution operation result filtering unit further comprises:
and the extension area information determining module is used for determining extension area information of the operation result size of the first convolution layer according to the size information needing to be extended, wherein the extension area information comprises column number information extending towards the right side and/or row number information extending towards the lower side.
17. The apparatus of claim 16, wherein the column value extending to the right is a remainder of a division of the number of columns of the actual screening area size by the step length s;
and the downward extension line numerical value is the remainder of division of the line number of the actual screening area size and the stepping length s.
18. The apparatus according to any one of claims 14 to 17, wherein the first convolution operation result screening unit is configured to:
the effective value screening module carries out non-overlapping screening on the first convolution layer operation result from top to bottom from left to right in sequence from the initial position in the first convolution layer operation result.
19. The apparatus according to claim 16 or 17, wherein the value in the extension area is set to an invalid value.
20. The apparatus of any of claims 11, 12, and 14-17, wherein the apparatus further comprises:
and the convolutional layer operation strategy information unit is used for presetting convolutional layer operation strategy information in the convolutional network, wherein the convolutional layer operation strategy information at least comprises convolutional layer operation size information and corresponding step length value information.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911041729.2A CN110826694B (en) | 2019-10-30 | 2019-10-30 | Image processing method and device based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911041729.2A CN110826694B (en) | 2019-10-30 | 2019-10-30 | Image processing method and device based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110826694A CN110826694A (en) | 2020-02-21 |
| CN110826694B true CN110826694B (en) | 2021-06-11 |
Family
ID=69551324
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911041729.2A Active CN110826694B (en) | 2019-10-30 | 2019-10-30 | Image processing method and device based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110826694B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113240077B (en) * | 2021-04-27 | 2022-04-05 | 瀚博半导体(上海)有限公司 | Tensor processing method and system |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109214308A (en) * | 2018-08-15 | 2019-01-15 | 武汉唯理科技有限公司 | A kind of traffic abnormity image identification method based on focal loss function |
| CN110309837A (en) * | 2019-07-05 | 2019-10-08 | 北京迈格威科技有限公司 | Data processing method and image processing method based on convolutional neural networks characteristic pattern |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017129325A1 (en) * | 2016-01-29 | 2017-08-03 | Fotonation Limited | A convolutional neural network |
| CN109426858B (en) * | 2017-08-29 | 2021-04-06 | 京东方科技集团股份有限公司 | Neural network, training method, image processing method, and image processing apparatus |
| CN107516304A (en) * | 2017-09-07 | 2017-12-26 | 广东工业大学 | A kind of image de-noising method and device |
| CN108510058B (en) * | 2018-02-28 | 2021-07-20 | 中国科学院计算技术研究所 | Weight storage method in neural network and processor based on method |
| CN110163042B (en) * | 2018-04-13 | 2023-05-30 | 腾讯科技（深圳）有限公司 | Image recognition method and device |
| CN110222829A (en) * | 2019-06-12 | 2019-09-10 | 北京字节跳动网络技术有限公司 | Feature extracting method, device, equipment and medium based on convolutional neural networks |
-
2019
- 2019-10-30 CN CN201911041729.2A patent/CN110826694B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109214308A (en) * | 2018-08-15 | 2019-01-15 | 武汉唯理科技有限公司 | A kind of traffic abnormity image identification method based on focal loss function |
| CN110309837A (en) * | 2019-07-05 | 2019-10-08 | 北京迈格威科技有限公司 | Data processing method and image processing method based on convolutional neural networks characteristic pattern |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110826694A (en) | 2020-02-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110555847B (en) | Image processing method and device based on convolutional neural network | |
| US11157592B2 (en) | Hardware implementation of convolutional layer of deep neural network | |
| CN108268869B (en) | Target detection method, device and system | |
| KR102631381B1 (en) | Convolutional neural network processing method and apparatus | |
| CN110494867B (en) | Electronic device for operating machine learning and method for operating machine learning | |
| CN112396115B (en) | Attention mechanism-based target detection method and device and computer equipment | |
| CN110598788B (en) | Target detection method, target detection device, electronic equipment and storage medium | |
| EP3637327B1 (en) | Computing device and method | |
| JP2021507345A (en) | Fusion of sparse kernels to approximate the complete kernel of convolutional neural networks | |
| CN110826694B (en) | Image processing method and device based on convolutional neural network | |
| CN111709415B (en) | Target detection method, device, computer equipment and storage medium | |
| JP4212676B2 (en) | Information processing system and information processing method | |
| JP3877002B2 (en) | Information processing system and information processing method | |
| CN111967478B (en) | Feature map reconstruction method, system, storage medium and terminal based on weight overturn | |
| CN114254731A (en) | Method and apparatus for neural network operation | |
| EP4307211A1 (en) | Method and apparatus with image restoration | |
| CN112764764A (en) | Scene model deployment method, device, equipment and storage medium | |
| CN108389153A (en) | A kind of method and terminal device of view load | |
| CN112766397A (en) | Classification network and implementation method and device thereof | |
| CN111860287A (en) | Target detection method and device and storage medium | |
| CN115865486B (en) | Network intrusion detection method and system based on multi-layer perception convolutional neural network | |
| CN101562691B (en) | Image processing device and method | |
| CN108769521A (en) | A kind of photographic method, mobile terminal and computer readable storage medium | |
| CN109961083B (en) | Method and image processing entity for applying a convolutional neural network to an image | |
| CN112215754A (en) | Image amplification method, device, electronic equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |