CN111783806A - Deep learning model optimization method and device and server - Google Patents
Deep learning model optimization method and device and server Download PDFInfo
- Publication number
- CN111783806A CN111783806A CN201910275218.0A CN201910275218A CN111783806A CN 111783806 A CN111783806 A CN 111783806A CN 201910275218 A CN201910275218 A CN 201910275218A CN 111783806 A CN111783806 A CN 111783806A
- Authority
- CN
- China
- Prior art keywords
- data
- learning model
- training
- optimization
- simulation
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Electrically Operated Instructional Devices (AREA)
Abstract
The invention is suitable for the technical field of positioning, and provides an optimization method, an optimization device and a server of a deep learning model, wherein the optimization method comprises the following steps: carrying out data augmentation based on an optimization instruction to obtain augmented data results, wherein the optimization instruction carries scene information; and optimizing the deep learning model based on the augmented data result to obtain an optimized learning model. In the invention, more data are obtained after the basic data are expanded, and then the optimization of the learning model is carried out, so that the optimization effect of the learning model can be enhanced.
Description
Technical Field
The invention belongs to the technical field of satellite positioning, and particularly relates to an optimization method and device of a deep learning model and a server.
Background
In recent years, with the continuous development of hardware computing power and the continuous upgrade of high-performance GPUs, the prospect of artificial intelligence technology becomes clear. And the computer vision field of deep learning is more prominent. Basic image processing algorithms such as face recognition, image classification, face detection, image segmentation and the like belong to the technology in the field of computer vision.
With the rapid development of the information era, the generation of mass data also lays a foundation for the rapid advance of deep learning. The most important factor influencing the good and bad effect of the deep learning model is data. Mass data represents the strong generalization ability of the model; conversely, lack of sufficient data can subject the model to over-or under-fitting problems.
With the rapid development of the deep learning technology, the model performance is continuously improved, and the application degree in the actual production is gradually increased. Deep learning model design and tuning for specific fields and specific problems are always very challenging problems. It is known that data is a key factor affecting final performance in the deep learning model research process. If the data is sufficient and the real situation can be well reflected, the performance of the trained model is generally good; if the data is lacking and does not reflect the real situation, the final model is easily problematic. It is a more effective means to adjust and optimize the deep learning model by adding real data, and in the prior art, the number of data is increased by adopting an original data expansion mode, but the data generated by the mode is more limited and cannot be really fused with the background, so that the learning model cannot be effectively optimized.
Disclosure of Invention
The embodiment of the invention provides an optimization method and device of a deep learning model and a server, and aims to solve the problem that the optimization effect of the learning model in the prior art is poor.
A method for optimizing a deep learning model comprises the following steps:
carrying out data augmentation based on an optimization instruction to obtain an augmented data result, wherein the optimization instruction carries scene information;
and optimizing the deep learning model based on the augmented data result to obtain an optimized learning model.
Preferably, the data augmentation is performed based on the optimization indication, and obtaining augmented data results comprises:
analyzing problems existing in a scene corresponding to the scene information to obtain an analysis result;
and performing data simulation based on the analysis result to obtain simulation data.
Preferably, the data simulation is performed based on the analysis result, and after obtaining the simulation data, the method further includes:
and carrying out data amplification based on the simulation data to obtain an amplified data result.
Preferably, performing data simulation based on the analysis result, and obtaining simulation data includes:
extracting characteristics of each of the existing problems based on the analysis results;
and performing corresponding data simulation based on the extracted features to obtain simulation data.
Preferably, the data augmentation is performed based on the simulation data, and obtaining augmented data results includes:
analyzing the simulation data to obtain an analysis result;
and performing data augmentation by adopting a corresponding augmentation rule according to a preset rule base and the analysis result to obtain an augmented data result.
Preferably, the deep learning model is optimized based on the augmented data result, and the optimized learning model is obtained by: and optimizing the deep learning model by adopting an iteration mode based on the augmented data result to obtain an optimized learning model.
Preferably, the optimization instruction further includes basic data, and the deep learning model is optimized in an iterative manner based on the augmented data result, and obtaining an optimized learning model includes:
training based on the basic data to obtain a first training model;
training the first training model based on the augmented data result to obtain a second training model;
and carrying out optimization analysis on the second training model to obtain an optimized learning model.
Preferably, training based on the basic data, and obtaining a first training model includes:
establishing a test set based on the basic data;
establishing a training set based on the test set and the simulation data;
and training based on the training set to obtain a first training model.
Preferably, the building a test set based on the base data comprises:
judging whether the quantity of the basic data is smaller than a preset value or not;
if so, establishing a test set based on the basic data and the simulation data;
and when the judgment is negative, establishing a test set based on the basic data.
Preferably, training the first training model based on the augmented data result, and obtaining a second training model comprises:
establishing a data set based on the training set and the augmented data;
and training based on the data set to obtain a second training model.
Preferably, performing optimization analysis on the second training model to obtain an optimized learning model includes:
testing the second training model on the data set to obtain a test result;
analyzing the test result, and performing corresponding data expansion on the training set according to the analysis result to obtain an expanded data set;
and training the second training model based on the expansion data set to obtain an optimized learning model.
Preferably, training the second training model based on the augmented data set, and obtaining the optimized learning model further includes:
testing the optimized learning model to obtain a testing effect;
analyzing whether the test effect meets a preset requirement or not;
and if not, turning to the step of updating the second training model based on the expansion data to obtain an optimized learning model.
The invention also provides an optimization device of the deep learning model, which comprises the following components:
the augmentation unit is used for augmenting data based on the optimization instruction to obtain augmented data results, and the optimization instruction carries scene information;
and the optimization unit is used for optimizing the deep learning model based on the augmented data result to obtain an optimized learning model.
The invention also provides a server, which comprises an optimization device of the deep learning model, wherein the optimization device comprises:
the augmentation unit is used for augmenting data based on the optimization instruction to obtain augmented data results, and the optimization instruction carries scene information;
and the optimization unit is used for optimizing the deep learning model based on the augmented data result to obtain an optimized learning model.
The invention also provides a memory storing a computer program executed by a processor to perform the steps of:
carrying out data augmentation based on an optimization instruction to obtain an augmented data result, wherein the optimization instruction carries scene information;
and optimizing the deep learning model based on the augmented data result to obtain an optimized learning model.
The invention also provides an optimization terminal, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to realize the following steps:
carrying out data augmentation based on an optimization instruction to obtain an augmented data result, wherein the optimization instruction carries scene information;
and optimizing the deep learning model based on the augmented data result to obtain an optimized learning model.
In the embodiment of the invention, more data are obtained after the basic data are augmented, and then the learning model is optimized, so that the optimization effect of the learning model can be enhanced.
Drawings
Fig. 1 is a flowchart of an optimization method of a deep learning model according to a first embodiment of the present invention;
fig. 2 is a flowchart illustrating a step S1 of a method for optimizing a deep learning model according to a first embodiment of the present invention;
fig. 3a is an actual schematic diagram of insulator string dropping in the optimization method of the deep learning model according to the first embodiment of the present invention;
fig. 3b is a schematic diagram of a simulation of insulator string dropping in the method for optimizing a deep learning model according to the first embodiment of the present invention;
fig. 3c is a schematic diagram of a simulation of insulator string dropping in the method for optimizing a deep learning model according to the first embodiment of the present invention;
fig. 3d is an actual schematic diagram of insulator damage of the optimization method of the deep learning model according to the first embodiment of the present invention;
fig. 3e is a schematic diagram of a hole shape gap of an insulator damage in the method for optimizing a deep learning model according to the first embodiment of the present invention;
fig. 4 is a flowchart illustrating a step S2 of a method for optimizing a deep learning model according to a first embodiment of the present invention;
FIG. 5 is a flowchart illustrating the step S21 of the method for optimizing a deep learning model according to the first embodiment of the present invention;
FIG. 6 is a block diagram of an optimization apparatus for deep learning model according to a second embodiment of the present invention;
fig. 7 is a block diagram of an optimized terminal according to a third embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
In an embodiment of the present invention, a method for optimizing a deep learning model includes: performing data amplification based on an optimization instruction to obtain an amplified data result, wherein the optimization instruction carries scene information; and optimizing the deep learning model based on the augmented data result to obtain an optimized learning model.
In order to explain the technical means of the present invention, the following description will be given by way of specific examples.
The first embodiment is as follows:
fig. 1 is a flowchart illustrating an optimization method of a deep learning model according to a first embodiment of the present invention, where the optimization method includes:
step S1, data augmentation is carried out based on the optimization instruction, and augmented data results are obtained;
specifically, data augmentation is performed according to the optimization instructions, resulting in augmented data results that include augmented data. The optimization instruction carries scene information and also carries basic data, data can be augmented according to problems which may occur in the scene, the basic data is real data in the scene, such as real problems existing in the scene, and the like, and preferably, the basic data may include problem data, such as pictures which are taken in the scene and reflect problems.
Step S2, optimizing the deep learning model based on the augmented data result to obtain an optimized learning model;
specifically, after data is augmented, the augmented data is input into a learning model for training and learning so as to optimize the learning model and obtain an optimized learning model.
In a preferable embodiment of this embodiment, before the step S1, the method further includes:
obtaining an optimization instruction;
specifically, an optimization instruction is obtained first, the optimization instruction can be sent by an administrator or a maintainer or a corresponding technician, the optimization instruction can include some basic data appearing in a certain scene, and because the amount of the basic data is small, expansion and augmentation of the data are needed to improve the optimization effect of the deep learning model.
In this embodiment, data expansion is performed on the basic data, and a large amount of data is obtained to optimize the deep learning model, so that the optimization effect can be enhanced.
In a preferred embodiment of this embodiment, as shown in fig. 2, a specific flowchart of step S1 of the method for optimizing a deep learning model according to the first embodiment of the present invention is provided, where the step S1 specifically includes:
step S11, analyzing the problems existing in the scene corresponding to the scene information to obtain an analysis result;
specifically, some problems that may occur in the corresponding scene are analyzed, data of real problems are analyzed, and types and distributions of the problems are classified, for example: under the scene of the electric power inspection project, for example, abnormal insulators need to be identified, and the existing problem types can include: insulator string falling, insulator damage and the like: each type of problem may in turn include a number of situations, such as: in the problem of insulator string dropping, the number, the positions and the like of the dropped strings are different correspondingly; the problem of insulator breakage includes the problem of the position, type and the like of breakage, and a corresponding analysis result is obtained. In this embodiment, some real problems collected on the web can be analyzed, and the insulator defect position is found to be analog. In the present embodiment, there is a problem of having simulatability.
Step S12, performing data simulation based on the analysis result to obtain simulation data;
specifically, according to the analysis result, corresponding data simulation is carried out aiming at the problem with the simulation property, and corresponding simulation data are obtained;
in a preferable scheme of this embodiment, the step S12 specifically includes:
extracting characteristics of the problems existing in each type based on the analysis result;
performing corresponding data simulation based on the extracted features to obtain simulation data;
specifically, the characteristics of each type of problem, which are simulative and generalizable, are analyzed and extracted, and then simulation is performed using image editing software to obtain simulation data, for example, manual simulation is performed on a normally shot picture to obtain simulation data;
for ease of understanding, the following description is given with a specific simulation example:
for the insulator string drop, as shown in fig. 3a, the insulator string drop is characterized in that one or more ceramic strings in a certain insulator drop, in a corresponding scene, the reason for causing the insulator string drop is usually that the insulator self-explosion is caused by overhigh voltage or the insulator drops due to hard object impact, such problems can not occur frequently, the acquisition of the corresponding data sample is also difficult, at this time, the data sample can be acquired in a manual simulation mode, and the simulation method is as follows: a painting tool, a color replacement tool and the like in the image editing software are adopted, and one or more ceramic strings in the normal insulator are manually removed from the image, so that the effect of simulating the string removal is achieved; controlling the falling number of the porcelain strings: analyzing the obtained small number of real insulator string dropping pictures (for example, from network search), and finding that the number of the ceramic strings dropping off is generally not more than 5 for a single insulator, so that the number of the ceramic strings dropping off is set to be a random value of 1-5 when insulator string dropping is artificially simulated; by the method, the data samples of the insulator string falling problem can be simulated and imitated manually, and the imitated samples can be used as insulator string falling simulation data. Examples of partial artificial simulation results are shown in FIGS. 3b and 3 c;
for insulator breakage:
as shown in fig. 3d, some or some of the porcelain strings in the insulator are partially damaged. In a real scene, the damage of the insulator is generally caused by the damage of a single or partial porcelain string due to the impact of a hard object or the careless collision in the installation process, and the like, but such problems do not occur frequently, and the acquisition of the corresponding data sample is difficult. At the moment, the formation of simulation data can be realized by adopting a manual simulation mode; preferably, the simulation is performed according to each type of feature:
for the triangular notch type (see fig. 3d), the characteristic is that a damaged notch appears locally on a single or partial ceramic string, the notch is shaped like a triangle, and the damaged proportion is the highest in the existing collected real data sample, so that the proportion can be increased when the damaged data is simulated, and the simulation mode can be: on a single or partial ceramic string of a normal insulator picture, a painting tool, a color replacement tool and the like in image editing software are adopted to artificially imitate the triangular gap, and the triangular shape can be diversified, such as an acute triangle, a right triangle and an obtuse triangle (namely, the gap on the ceramic string can be an acute angle, a right angle and an obtuse angle);
for the hole shape gap type (see fig. 3e), the method is characterized in that a single or partial ceramic string has a local damaged gap, and the gap is similar to a hole in shape; the notch of the hole shape can be artificially imitated on a single or partial ceramic string of a normal insulator picture by adopting a painting tool, a color replacing tool and the like in image editing software. Wherein the shape of the holes can be varied, such as circular, semicircular, elliptical, semi-elliptical, etc.;
for linear notch types: is characterized in that a damaged gap appears on the local part of a single or partial ceramic string, and the shape of the gap is similar to a linear shape; the linear gap can be artificially imitated on a single or partial ceramic string of a normal insulator picture by adopting a painting tool, a color replacing tool and the like in image editing software. The cutting degree (namely the length of the straight line) of the linear notch can be diversified.
For irregular shaped notch types: is characterized in that the local part of a single or partial porcelain string has a damaged gap, and the gap is in an irregular shape, such as a sawtooth shape and the like; the irregular-shaped gap can be artificially imitated on a single or partial ceramic string of a normal insulator picture by using a painting tool, a color replacement tool and the like in image editing software;
for crack shape notch types: the porcelain string is characterized in that obvious cracks appear on the local part of a single or partial porcelain string, and the cracks comprise linear cracks, curved cracks and the like. Such cracks can be artificially imitated on a single or partial ceramic string of a normal insulator picture by using a painting tool, a color replacement tool, and the like in image editing software. Wherein the shape of the cracks can be linear or curvilinear.
By simulating the simulation data of the sample with the problem of insulator breakage in the manual simulation mode, the problem of sample data shortage can be effectively solved.
In a preferable embodiment of this embodiment, after step S12, the method further includes:
step S13, performing data augmentation based on the simulation data to obtain augmented data results;
specifically, the analog data is augmented to obtain augmented data, that is, the augmented data is augmented from the analog data of the above-mentioned manual simulation, and the augmented data is automatically generated by using techniques such as image processing and deep learning.
In a preferable scheme of this embodiment, the step S13 specifically includes:
analyzing the simulation data to obtain an analysis result;
performing data augmentation by adopting a corresponding augmentation rule according to a preset rule base and the analysis result to obtain an augmented data result;
specifically, a set of complete rule base for automatic data augmentation is preset, and the rule base includes 17 augmentation rules, specifically: 15 image processing methods and 2 methods based on deep learning data generation, which can be used for amplifying simulation data, wherein the 15 image processing methods comprise: horizontal mirror image, vertical mirror image, image rotation, image enhancement, color attenuation, brightness pair comparison adjustment, mean value filtering, Gaussian filtering, median filtering, bilateral filtering, salt and pepper noise, Gaussian noise, image random clipping, image random translation and image random miscut, wherein the 2 methods based on deep learning data generation respectively comprise the following steps: generating a confrontation network method and an image style migration method; the simulation effect corresponding to each of the aforementioned augmentation rules is shown in table 1;
further, the simulation data is analyzed, for example, the types of problems (such as image blur, light darkness, image rotation, etc.) that may occur in the real scene are analyzed to obtain corresponding analysis results, and then, according to the analysis results and the rule base, data augmentation is performed by using corresponding augmentation rules to obtain augmented data results, preferably, the characteristics of the real scene need to be analyzed, and then, the augmentation rules that need to be used are selected by combining table 1.
TABLE 1
In an actual application scenario, taking an insulator as an example, an original picture which is taken at the beginning is taken by an unmanned aerial vehicle, a flight angle of the unmanned aerial vehicle is arbitrary, the same insulator may have a plurality of different postures, which are also arbitrary in an image, and an augmentation rule which needs to be adopted at this time is as follows: horizontal mirroring, vertical mirroring, image rotation, image random cutting and image random translation; in addition, unmanned aerial vehicle is at the flight in-process, very easy emergence shake, and flight is unstable, causes the image of shooing to appear that picture is fuzzy, luminance contrast variation, colour distortion, produce noise scheduling problem easily, and the augmentation rule of adoption includes: brightness contrast adjustment, color attenuation, image enhancement, mean filtering, gaussian filtering, median filtering, bilateral filtering, salt-pepper noise, gaussian noise and the like, so that 14 kinds of augmentation rules are required to be adopted for data augmentation, for example, 1200 pictures are included in analog data for data augmentation to obtain 18000 pictures, and therefore the problem of data shortage is solved.
In a preferable scheme of this embodiment, the step S2 specifically includes: optimizing the deep learning model by adopting an iteration mode based on the augmented data result to obtain an optimized learning model;
further, as shown in fig. 4, a detailed flowchart of step S2 of the method for optimizing a deep learning model according to the first embodiment of the present invention is provided, where the step S2 specifically includes:
step S21, performing first training based on basic data to obtain a first training model;
specifically, training is performed based on basic data to obtain a first training model;
step S22, training the first training model based on the augmented data result to obtain a second training model;
specifically, a first training model is trained based on the augmented data result to obtain a second training model;
step S23, carrying out optimization analysis on the second training model to obtain an optimized learning model;
specifically, the second training model is further optimized and analyzed to obtain an optimized learning model;
in a further preferred embodiment of this embodiment, as shown in fig. 5, a specific flowchart of step S21 of the method for optimizing a deep learning model according to the first embodiment of the present invention is provided, where the step S21 specifically includes:
step S211, establishing a test set based on the basic data;
specifically, a test set is established according to basic data, in this embodiment, the number of the basic data may be relatively small, and the basic data is used as the test set;
in a preferred embodiment of the present invention, if the amount of the basic data is too small or not, part of the simulation data may be added as the test set, and the amount of the added simulation data may be determined according to the actual situation, which is not limited herein.
Step S212, establishing a training set based on the test set and the simulation data;
specifically, a training set is established based on the aforementioned test set;
step S213, training is carried out based on the training set to obtain a first training model;
in a further preferable embodiment of this embodiment, after step S213, the method may further include:
testing the effect of the first training model on the test set to obtain a corresponding test effect;
in a further preferable scheme of this embodiment, the step S212 specifically includes:
judging whether the quantity of the basic data is smaller than a preset value or not;
if so, establishing a test set based on the basic data and the simulation data;
when the judgment is negative, establishing a test set based on the basic data;
specifically, it is to be determined whether the amount of the basic data in the test set is smaller than a preset value, and when the amount of the basic data is not smaller than the preset value, it is indicated that the amount of the basic data is sufficient, at this time, the test set may be established according to the basic data, that is, the basic data is used as a training set, and when the amount of the basic data is smaller than the preset value, it is indicated that the basic data is insufficient, at this time, a part or all of the simulation data and the basic data need to be combined to be used as the training set, it is to be indicated that a part of the simulation data or all of the simulation data may be selected according to an actual situation, which is not limited herein.
In a variation of this embodiment, the base data may be 0, that is, there is no data available for building the test set in the base data, and at this time, part or all of the simulation data is also selected as the training set, and the data in the test set at this time is all from the simulation data.
In another variation of this embodiment, if the sum of the quantities of the simulation data and the basic data is less than a threshold, a part of the augmented data may be added as the training set, and the specific size of the threshold may be set according to the actual situation, which is not limited herein, and the quantity of the augmented data added to the training set may also be determined according to the actual situation, which is also not limited herein.
In a preferable scheme of this embodiment, the step S22 specifically includes:
establishing a data set based on the training set and the augmented data;
specifically, firstly, combining the data of the training set and the augmented data to form a data set;
training based on the data set to obtain a second training model;
specifically, the data set is used for training a first training model to obtain a corresponding second training model;
in a preferable scheme of this embodiment, the step S23 specifically includes:
testing the second training model on the data set to obtain a test result;
analyzing the test result, and performing corresponding data expansion on the data set according to the analysis result to obtain an expanded data set;
specifically, the effect of the second training model is tested on the data set to obtain a test result, the test result is analyzed to determine problems of the second training model (for example, which characteristics of a picture with an error in identification on the data set are fuzzy, dark light and the like), and then corresponding data expansion is performed according to the existing problems, namely, the data of the data set is expanded in a targeted manner, and the effect of the picture with the problems is simulated to expand to obtain an expanded data set;
training the second training model based on the extended data set to obtain an optimized learning model;
in a further preferred embodiment of this embodiment, the step of training the second training model based on the extended data set further includes, after obtaining the optimized learning model:
testing the optimized learning model to obtain a testing effect;
analyzing whether the test effect meets the preset requirement or not;
if the judgment result is no, training a second training model based on the expansion data set to obtain an optimized learning model, and if the judgment result is yes, stopping the process;
in a preferred embodiment of the present invention, the preset requirement may be set according to an actual situation, for example, the preset requirement may be specific parameters (such as pixel values, resolution, etc.) of a picture;
it should be noted that the first training model and the second training model have no special meaning such as sequence, and are only used for distinction and explanation.
In this embodiment, the learning training is performed through less data, the test effect of the obtained learning model is tested, data expansion is performed according to the test effect in a targeted manner, the learning model is optimized step by step, and the optimization efficiency of the learning model can be improved.
In a preferable embodiment of this embodiment, after step S2, the method further includes:
accumulating scene data based on the optimized learning model;
specifically, an optimized learning model is put into a current scene for use, basic data corresponding to the current scene are collected, the optimized learning model is used for carrying out test simulation on the basic data, problem data obtained from a test are screened, the problem data left after screening are added into the basic data for accumulation, and when the preset number of the problem data is accumulated, the problem data can be added into a training set so as to further optimize the learning model in the follow-up process.
In the embodiment, the basic data are augmented to obtain more data, and then the learning model is optimized, so that the optimization effect of the learning model can be enhanced;
secondly, firstly, learning training is carried out through less data, the testing effect of the obtained learning model is tested, data expansion is carried out according to the testing effect in a targeted mode, the learning model is optimized step by step, and the optimization efficiency of the learning model can be improved.
Example two:
based on the first embodiment, as shown in fig. 6, a structural diagram of an optimization apparatus for a depth learning model according to a second embodiment of the present invention is provided, where the optimization apparatus includes: augmentation unit 1 and optimization unit 2 connected with it, wherein:
the augmentation unit 1 is used for performing data augmentation based on the optimization instruction to obtain augmented data results;
specifically, data augmentation is performed according to the optimization instructions, resulting in augmented data results that include augmented data. The optimization instruction carries scene information and also carries basic data, data can be augmented according to problems which may occur in the scene, the basic data is real data in the scene, such as real problems existing in the scene, and the like, and preferably, the basic data may include problem data, such as pictures which are taken in the scene and reflect problems.
The optimization unit 2 is used for optimizing the deep learning model based on the augmented data result to obtain an optimized learning model;
specifically, after data is augmented, the augmented data is input into a learning model for training and learning so as to optimize the learning model and obtain an optimized learning model.
In a preferable aspect of this embodiment, the optimizing device may further include: an acquisition unit connected to the augmentation unit 1, wherein:
an obtaining unit configured to obtain an optimization instruction;
specifically, an optimization instruction is obtained first, the optimization instruction can be sent by an administrator or a maintainer or a corresponding technician, the optimization instruction can include some basic data appearing in a certain scene, and because the amount of the basic data is small, expansion and augmentation of the data are needed to improve the optimization effect of the deep learning model.
In this embodiment, data expansion is performed on the basic data, and a large amount of data is obtained to optimize the deep learning model, so that the optimization effect can be enhanced.
In a preferred embodiment of this embodiment, the amplification unit 1 specifically includes: an analysis subunit, a simulation subunit connected to the analysis subunit, wherein:
the analysis subunit is used for analyzing the problems in the scene corresponding to the scene information to obtain an analysis result;
specifically, some problems that may occur in the corresponding scene are analyzed, data of real problems are analyzed, and types and distributions of the problems are classified, for example: under the scene of the electric power inspection project, for example, abnormal insulators need to be identified, and the existing problem types can include: insulator string falling, insulator damage and the like: each type of problem may in turn include a number of situations, such as: in the problem of insulator string dropping, the number, the positions and the like of the dropped strings are different correspondingly; the problem of insulator breakage includes the problem of the position, type and the like of breakage, and a corresponding analysis result is obtained. In this embodiment, some real problems collected on the web can be analyzed, and the insulator defect position is found to be analog. In the present embodiment, there is a problem of having simulatability.
The simulation subunit is used for carrying out data simulation based on the analysis result to obtain simulation data;
specifically, according to the analysis result, corresponding data simulation is carried out aiming at the problem with the simulation property, and corresponding simulation data are obtained;
the analog subunit is specifically configured to:
extracting characteristics of the problems existing in each type based on the analysis result;
performing corresponding data simulation based on the extracted features to obtain simulation data;
specifically, the characteristics of each type of problem, which are simulative and generalizable, are analyzed and extracted, and then simulation is performed using image editing software to obtain simulation data, for example, manual simulation is performed on a normally shot picture to obtain simulation data;
for ease of understanding, the following description is given with a specific simulation example:
for the insulator string drop, as shown in fig. 3a, the insulator string drop is characterized in that one or more ceramic strings in a certain insulator drop, in a corresponding scene, the reason for causing the insulator string drop is usually that the insulator self-explosion is caused by overhigh voltage or the insulator drops due to hard object impact, such problems can not occur frequently, the acquisition of the corresponding data sample is also difficult, at this time, the data sample can be acquired in a manual simulation mode, and the simulation method is as follows: a painting tool, a color replacement tool and the like in the image editing software are adopted, and one or more ceramic strings in the normal insulator are manually removed from the image, so that the effect of simulating the string removal is achieved; controlling the falling number of the porcelain strings: analyzing the obtained small number of real insulator string dropping pictures (for example, from network search), and finding that the number of the ceramic strings dropping off is generally not more than 5 for a single insulator, so that the number of the ceramic strings dropping off is set to be a random value of 1-5 when insulator string dropping is artificially simulated; by the method, the data samples of the insulator string falling problem can be simulated and imitated manually, and the imitated samples can be used as insulator string falling simulation data. Examples of partial artificial simulation results are shown in FIGS. 3b and 3 c;
for insulator breakage:
as shown in fig. 3d, some or some of the porcelain strings in the insulator are partially damaged. In a real scene, the damage of the insulator is generally caused by the damage of a single or partial porcelain string due to the impact of a hard object or the careless collision in the installation process, and the like, but such problems do not occur frequently, and the acquisition of the corresponding data sample is difficult. At the moment, the formation of simulation data can be realized by adopting a manual simulation mode; preferably, the simulation is performed according to each type of feature:
for the triangular notch type (see fig. 3e), the characteristic is that a damaged notch appears locally on a single or partial ceramic string, the notch is shaped like a triangle, and the damaged proportion is the highest in the existing collected real data sample, so that the proportion can be increased when the damaged data is simulated, and the simulation mode can be: on a single or partial ceramic string of a normal insulator picture, a painting tool, a color replacement tool and the like in image editing software are adopted to artificially imitate the triangular gap, and the triangular shape can be diversified, such as an acute triangle, a right triangle and an obtuse triangle (namely, the gap on the ceramic string can be an acute angle, a right angle and an obtuse angle);
for the hole shape gap type (see fig. 3e), the method is characterized in that a single or partial ceramic string has a local damaged gap, and the gap is similar to a hole in shape; the notch of the hole shape can be artificially imitated on a single or partial ceramic string of a normal insulator picture by adopting a painting tool, a color replacing tool and the like in image editing software. Wherein the shape of the holes can be varied, such as circular, semicircular, elliptical, semi-elliptical, etc.;
for linear notch types: is characterized in that a damaged gap appears on the local part of a single or partial ceramic string, and the shape of the gap is similar to a linear shape; the linear gap can be artificially imitated on a single or partial ceramic string of a normal insulator picture by adopting a painting tool, a color replacing tool and the like in image editing software. The cutting degree (namely the length of the straight line) of the linear notch can be diversified.
For irregular shaped notch types: is characterized in that the local part of a single or partial porcelain string has a damaged gap, and the gap is in an irregular shape, such as a sawtooth shape and the like; the irregular-shaped gap can be artificially imitated on a single or partial ceramic string of a normal insulator picture by using a painting tool, a color replacement tool and the like in image editing software;
for crack shape notch types: the porcelain string is characterized in that obvious cracks appear on the local part of a single or partial porcelain string, and the cracks comprise linear cracks, curved cracks and the like. Such cracks can be artificially imitated on a single or partial ceramic string of a normal insulator picture by using a painting tool, a color replacement tool, and the like in image editing software. Wherein the shape of the cracks can be linear or curvilinear.
By simulating the simulation data of the sample with the problem of insulator breakage in the manual simulation mode, the problem of sample data shortage can be effectively solved.
In a preferable aspect of the present embodiment, the amplification unit 1 further includes: an amplification subunit connected to the analog subunit, wherein:
the amplification subunit is used for carrying out data amplification based on the analog data to obtain an amplified data result;
specifically, the analog data is augmented to obtain augmented data, that is, the augmented data is augmented from the analog data of the above-mentioned manual simulation, and the augmented data is automatically generated by using techniques such as image processing and deep learning.
The amplification subunit is specifically configured to:
analyzing the simulation data to obtain an analysis result;
performing data augmentation by adopting a corresponding augmentation rule according to a preset rule base and the analysis result to obtain an augmented data result;
specifically, a set of complete rule base for automatic data augmentation is preset, and the rule base includes 17 augmentation rules, specifically: 15 image processing methods and 2 methods based on deep learning data generation, which can be used for amplifying simulation data, and as shown in table 1, the simulation effect corresponding to each of the aforementioned augmentation rules is shown;
further, the simulation data is analyzed, for example, the types of problems (such as image blur, light darkness, image rotation, etc.) that may occur in the real scene are analyzed to obtain corresponding analysis results, and then, according to the analysis results and the rule base, data augmentation is performed by using corresponding augmentation rules to obtain augmented data results, preferably, the characteristics of the real scene need to be analyzed, and then, the augmentation rules that need to be used are selected by combining table 1.
In an actual application scenario, taking an insulator as an example, an original picture which is taken at the beginning is taken by an unmanned aerial vehicle, a flight angle of the unmanned aerial vehicle is arbitrary, the same insulator may have a plurality of different postures, which are also arbitrary in an image, and an augmentation rule which needs to be adopted at this time is as follows: horizontal mirroring, vertical mirroring, image rotation, image random cutting and image random translation; in addition, unmanned aerial vehicle is at the flight in-process, very easy emergence shake, and flight is unstable, causes the image of shooing to appear that picture is fuzzy, luminance contrast variation, colour distortion, produce noise scheduling problem easily, and the augmentation rule of adoption includes: brightness contrast adjustment, color attenuation, image enhancement, mean filtering, gaussian filtering, median filtering, bilateral filtering, salt-pepper noise, gaussian noise and the like, so that 14 kinds of augmentation rules are required to be adopted for data augmentation, for example, 1200 pictures are included in analog data for data augmentation to obtain 18000 pictures, and therefore the problem of data shortage is solved.
In a preferred embodiment of this embodiment, the optimizing unit 2 specifically includes: a first training subunit, a second training subunit connected with the first training subunit, and an optimization subunit connected with the second training subunit, wherein:
the first training subunit is used for carrying out first training based on the basic data to obtain a first training model;
specifically, training is performed based on basic data to obtain a first training model;
the second training subunit is used for training the first training model based on the augmented data result to obtain a second training model;
specifically, a first training model is trained based on the augmented data result to obtain a second training model;
the optimization subunit is used for carrying out optimization analysis on the second training model to obtain an optimized learning model;
specifically, the second training model is further optimized and analyzed to obtain an optimized learning model;
in a further preferred embodiment of this embodiment, the first training subunit is specifically configured to:
firstly, establishing a test set based on basic data;
specifically, a test set is established according to basic data, in this embodiment, the number of the basic data may be relatively small, and the basic data is used as the test set;
in a preferred embodiment of the present invention, if the amount of the basic data is too small or not, part of the simulation data may be added as the test set, and the amount of the added simulation data may be determined according to the actual situation, which is not limited herein.
Secondly, establishing a training set based on the test set and the simulation data;
specifically, a training set is established based on the aforementioned test set;
then, training is carried out based on the training set to obtain a first training model;
in a further preferred aspect of this embodiment, the first training subunit is further configured to:
testing the effect of the first training model on the test set to obtain a corresponding test effect;
in a further preferred embodiment of this embodiment, the process of establishing the training set based on the test set and the simulation data is as follows:
judging whether the quantity of the basic data is smaller than a preset value or not;
if so, establishing a test set based on the basic data and the simulation data;
when the judgment is negative, establishing a test set based on the basic data;
specifically, it is to be determined whether the amount of the basic data in the test set is smaller than a preset value, and when the amount of the basic data is not smaller than the preset value, it is indicated that the amount of the basic data is sufficient, at this time, the test set may be established according to the basic data, that is, the basic data is used as a training set, and when the amount of the basic data is smaller than the preset value, it is indicated that the basic data is insufficient, at this time, a part or all of the simulation data and the basic data need to be combined to be used as the training set, it is to be indicated that a part of the simulation data or all of the simulation data may be selected according to an actual situation, which is not limited herein.
In a variation of this embodiment, the base data may be 0, that is, there is no data available for building the test set in the base data, and at this time, part or all of the simulation data is also selected as the training set, and the data in the test set at this time is all from the simulation data.
In a variation of this embodiment, the base data may be 0, that is, there is no data available for building the test set in the base data, and at this time, part or all of the simulation data is also selected as the training set, and the data in the test set at this time is all from the simulation data.
In a preferred embodiment of this embodiment, the second training subunit is specifically configured to:
firstly, establishing a data set based on a training set and augmented data;
specifically, firstly, combining the data of the training set and the augmented data to form a data set;
secondly, training based on the data set to obtain a second training model;
specifically, the data set is used for training a first training model to obtain a corresponding second training model;
in a preferred embodiment of this embodiment, the optimization subunit is specifically configured to:
testing the second training model on the data set to obtain a test result;
analyzing the test result, and performing corresponding data expansion on the data set according to the analysis result to obtain an expanded data set;
specifically, the effect of the second training model is tested on the data set to obtain a test result, the test result is analyzed to determine problems of the second training model (for example, which characteristics of a picture with an error in identification on the data set are fuzzy, dark light and the like), and then corresponding data expansion is performed according to the existing problems, namely, the data of the data set is expanded in a targeted manner, and the effect of the picture with the problems is simulated to expand to obtain an expanded data set;
training the second training model based on the extended data set to obtain an optimized learning model;
in a further preferred embodiment of this embodiment, the optimizing subunit is further configured to:
testing the optimized learning model to obtain a testing effect;
analyzing whether the test effect meets the preset requirement or not;
if the judgment result is no, training a second training model based on the expansion data set to obtain an optimized learning model, and if the judgment result is yes, stopping the process;
in a preferred embodiment of the present invention, the preset requirement may be set according to an actual situation, for example, the preset requirement may be specific parameters (such as pixel values, resolution, etc.) of a picture;
it should be noted that the first training model and the second training model have no special meaning such as sequence, and are only used for distinction and explanation.
In this embodiment, the learning training is performed through less data, the test effect of the obtained learning model is tested, data expansion is performed according to the test effect in a targeted manner, the learning model is optimized step by step, and the optimization efficiency of the learning model can be improved.
In a preferable aspect of this embodiment, the optimizing device further includes: an accumulation unit connected to the optimization unit 2, wherein:
an accumulation unit configured to accumulate scene data based on the optimized learning model;
specifically, an optimized learning model is put into a current scene for use, basic data corresponding to the current scene are collected, the optimized learning model is used for carrying out test simulation on the basic data, problem data obtained from a test are screened, the problem data left after screening are added into the basic data for accumulation, and when the preset number of the problem data is accumulated, the problem data can be added into a training set so as to further optimize the learning model in the follow-up process.
In the embodiment, the basic data are augmented to obtain more data, and then the learning model is optimized, so that the optimization effect of the learning model can be enhanced;
secondly, firstly, learning training is carried out through less data, the testing effect of the obtained learning model is tested, data expansion is carried out according to the testing effect in a targeted mode, the learning model is optimized step by step, and the optimization efficiency of the learning model can be improved.
In the present invention, a server is further provided, where the server includes an optimization device of a deep learning model, and the specific structure, the working principle, and the technical effects of the optimization device are consistent with the descriptions of the second embodiment, and are not described herein again.
Example three:
fig. 7 shows a block diagram of an optimized terminal according to a third embodiment of the present invention, where the optimized terminal includes: a memory (memory)71, a processor (processor)72, a communication Interface (Communications Interface)73 and a bus 74, wherein the processor 72, the memory 71 and the communication Interface 73 are in mutual communication via the bus 74.
A memory 71 for storing various data;
specifically, the memory 71 is used for storing various data, such as data in communication, received data, and the like, and is not limited herein, and the memory also includes a plurality of computer programs.
A communication interface 73 for information transmission between communication devices of the positioning terminal;
the processor 72 is used for calling various computer programs in the memory 71 to execute the optimization method of the deep learning model provided in the first embodiment, for example:
carrying out data augmentation based on an optimization instruction to obtain an augmented data result, wherein the optimization instruction carries scene information;
and optimizing the deep learning model based on the augmented data result to obtain an optimized learning model.
In the embodiment, more data are obtained after the basic data are augmented, and then the learning model is optimized, so that the optimization effect of the learning model can be enhanced;
the invention also provides a memory, wherein the memory stores a plurality of computer programs, and the computer programs are called by the processor to execute the optimization method of the deep learning model in the first embodiment.
In the invention, more data are obtained after the basic data are augmented, and then the optimization of the learning model is carried out, so that the optimization effect of the learning model can be enhanced;
secondly, firstly, learning training is carried out through less data, the testing effect of the obtained learning model is tested, data expansion is carried out according to the testing effect in a targeted mode, the learning model is optimized step by step, and the optimization efficiency of the learning model can be improved.
Those of ordinary skill in the art would appreciate that the elements and algorithm steps of the various embodiments described in connection with the embodiments disclosed herein may be implemented as electronic hardware, or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation.
Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention. The above description is only for the specific embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (16)
1. A method for optimizing a deep learning model, comprising:
carrying out data augmentation based on an optimization instruction to obtain augmented data results, wherein the optimization instruction carries scene information;
and optimizing the deep learning model based on the augmented data result to obtain an optimized learning model.
2. The optimization method of claim 1, wherein performing data augmentation based on the optimization indication, and obtaining augmented data results comprises:
analyzing problems existing in a scene corresponding to the scene information to obtain an analysis result;
and performing data simulation based on the analysis result to obtain simulation data.
3. The optimization method of claim 2, wherein performing data simulation based on the analysis result further comprises, after obtaining the simulation data:
and carrying out data amplification based on the simulation data to obtain an amplified data result.
4. The optimization method of claim 3, wherein performing data simulation based on the analysis result, and obtaining simulation data comprises:
extracting characteristics of each of the existing problems based on the analysis results;
and performing corresponding data simulation based on the extracted features to obtain simulation data.
5. The optimization method of claim 4, wherein performing data augmentation based on the simulation data, the obtaining augmented data results comprising:
analyzing the simulation data to obtain an analysis result;
and performing data augmentation by adopting a corresponding augmentation rule according to a preset rule base and the analysis result to obtain an augmented data result.
6. The optimization method according to claim 3, wherein the deep learning model is optimized based on the augmented data result, and the optimized learning model is specifically: and optimizing the deep learning model by adopting an iteration mode based on the augmented data result to obtain an optimized learning model.
7. The optimization method of claim 4, wherein the optimization instructions further include basic data, and the deep learning model is optimized in an iterative manner based on the augmented data results, and obtaining an optimized learning model includes:
training based on the basic data to obtain a first training model;
training the first training model based on the augmented data result to obtain a second training model;
and carrying out optimization analysis on the second training model to obtain an optimized learning model.
8. The optimization method of claim 5, wherein training based on the base data, resulting in a first training model comprises:
establishing a test set based on the basic data;
establishing a training set based on the test set and the simulation data;
and training based on the training set to obtain a first training model.
9. The optimization method of claim 8, wherein building a test set based on the base data comprises:
judging whether the quantity of the basic data is smaller than a preset value or not;
if so, establishing a test set based on the basic data and the simulation data;
and when the judgment is negative, establishing a test set based on the basic data.
10. The optimization method of claim 7, wherein training the first training model based on the augmented data results, resulting in a second training model comprises:
establishing a data set based on the training set and the augmented data;
and training based on the data set to obtain a second training model.
11. The optimization method of claim 10, wherein performing optimization analysis on the second training model to obtain an optimized learning model comprises:
testing the second training model on the data set to obtain a test result;
analyzing the test result, and performing corresponding data expansion on the training set according to the analysis result to obtain an expanded data set;
and training the second training model based on the expansion data set to obtain an optimized learning model.
12. The optimization method of claim 11, wherein training the second training model based on the augmented data set further comprises, after obtaining the optimized learning model:
testing the optimized learning model to obtain a testing effect;
analyzing whether the test effect meets a preset requirement or not;
and if not, turning to the step of updating the second training model based on the expansion data to obtain an optimized learning model.
13. An apparatus for optimizing a deep learning model, comprising:
the augmentation unit is used for augmenting data based on the optimization instruction to obtain augmented data results, and the optimization instruction carries scene information;
and the optimization unit is used for optimizing the deep learning model based on the augmented data result to obtain an optimized learning model.
14. A server, characterized by comprising means for optimizing a deep learning model according to claim 12.
15. A memory storing a computer program, the computer program being executable by a processor to perform the steps of:
carrying out data augmentation based on an optimization instruction to obtain augmented data results, wherein the optimization instruction carries scene information;
and optimizing the deep learning model based on the augmented data result to obtain an optimized learning model.
16. A data terminal comprising a memory, a processor and a computer program stored in said memory and executable on said processor, characterized in that said processor, when executing said computer program, implements the steps of the method for optimizing a deep learning model according to any one of claims 1 to 12.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910275218.0A CN111783806A (en) | 2019-04-04 | 2019-04-04 | Deep learning model optimization method and device and server |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910275218.0A CN111783806A (en) | 2019-04-04 | 2019-04-04 | Deep learning model optimization method and device and server |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111783806A true CN111783806A (en) | 2020-10-16 |
Family
ID=72755163
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910275218.0A Pending CN111783806A (en) | 2019-04-04 | 2019-04-04 | Deep learning model optimization method and device and server |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111783806A (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160293037A1 (en) * | 2015-03-30 | 2016-10-06 | Cae Inc. | Method and system for generating an interactive training scenario based on a recorded real time simulation |
| CN106875450A (en) * | 2017-02-20 | 2017-06-20 | 清华大学 | For the training set optimization method and device of camera reorientation |
| CN107145846A (en) * | 2017-04-26 | 2017-09-08 | 贵州电网有限责任公司输电运行检修分公司 | A kind of insulator recognition methods based on deep learning |
| CN107316004A (en) * | 2017-06-06 | 2017-11-03 | 西北工业大学 | Space Target Recognition based on deep learning |
| CN108107716A (en) * | 2017-12-19 | 2018-06-01 | 电子科技大学 | A kind of Parameter Measuring method based on improved BP neural network |
| CN109146880A (en) * | 2018-09-30 | 2019-01-04 | 亿嘉和科技股份有限公司 | A kind of electric device maintenance method based on deep learning |
| CN109190580A (en) * | 2018-09-15 | 2019-01-11 | 武汉智觉空间信息技术有限公司 | A kind of power components recognition methods of hybrid virtual scene assisted learning |
-
2019
- 2019-04-04 CN CN201910275218.0A patent/CN111783806A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160293037A1 (en) * | 2015-03-30 | 2016-10-06 | Cae Inc. | Method and system for generating an interactive training scenario based on a recorded real time simulation |
| CN106875450A (en) * | 2017-02-20 | 2017-06-20 | 清华大学 | For the training set optimization method and device of camera reorientation |
| CN107145846A (en) * | 2017-04-26 | 2017-09-08 | 贵州电网有限责任公司输电运行检修分公司 | A kind of insulator recognition methods based on deep learning |
| CN107316004A (en) * | 2017-06-06 | 2017-11-03 | 西北工业大学 | Space Target Recognition based on deep learning |
| CN108107716A (en) * | 2017-12-19 | 2018-06-01 | 电子科技大学 | A kind of Parameter Measuring method based on improved BP neural network |
| CN109190580A (en) * | 2018-09-15 | 2019-01-11 | 武汉智觉空间信息技术有限公司 | A kind of power components recognition methods of hybrid virtual scene assisted learning |
| CN109146880A (en) * | 2018-09-30 | 2019-01-04 | 亿嘉和科技股份有限公司 | A kind of electric device maintenance method based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| 欧啸天;胡伟;: "自然场景中Logo标识检测训练数据生成方法研究", 电子设计工程, no. 07 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110379020B (en) | Laser point cloud coloring method and device based on generation countermeasure network | |
| KR101891225B1 (en) | Method and apparatus for updating a background model | |
| CN109902018B (en) | Method for acquiring test case of intelligent driving system | |
| CN109472193A (en) | Method for detecting human face and device | |
| CN113505854B (en) | Face image quality evaluation model construction method, device, equipment and medium | |
| CN110348390B (en) | Training method, computer readable medium and system for flame detection model | |
| CN105812618B (en) | A kind of method for testing motion and motion detection apparatus | |
| CN105741244A (en) | Method for removing shadows and halos under weak light through indoor polling robot | |
| CN112036242B (en) | Face picture acquisition method and device, computer equipment and storage medium | |
| CN105678245A (en) | Target position identification method based on Haar features | |
| CN110807738A (en) | Fuzzy image non-blind restoration method based on edge image block sharpening | |
| CN108427971A (en) | The method and system of tobacco leaf grading based on mobile terminal | |
| CN113505702A (en) | Pavement disease identification method and system based on double neural network optimization | |
| CN116385758A (en) | Detection method for damage to surface of conveyor belt based on YOLOv5 network | |
| CN108765463A (en) | A kind of moving target detecting method calmodulin binding domain CaM extraction and improve textural characteristics | |
| CN115797314A (en) | Part surface defect detection method, system, equipment and storage medium | |
| CN111369477A (en) | Method for pre-analysis and tool self-adaptation of video recovery task | |
| Khan et al. | A deep hybrid few shot divide and glow method for ill-light image enhancement | |
| CN111783806A (en) | Deep learning model optimization method and device and server | |
| CN117409083A (en) | Cable terminal identification method and device based on infrared image and improved YOLOV5 | |
| CN111340839A (en) | Video noise estimation method based on block | |
| CN115294440A (en) | Power transmission line identification method and device and electronic equipment | |
| CN111738961B (en) | Electronic image restoration method, device and equipment and readable storage medium | |
| CN108010054B (en) | Method and system for extracting moving target of video image of segmented Gaussian mixture model | |
| CN110728686A (en) | Voronoi-based vehicle-mounted lamp image segmentation method |
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 |