CN110852352B - Data enhancement method for training deep neural network model for target detection - Google Patents

Data enhancement method for training deep neural network model for target detection Download PDF

Info

Publication number
CN110852352B
CN110852352B CN201911007178.8A CN201911007178A CN110852352B CN 110852352 B CN110852352 B CN 110852352B CN 201911007178 A CN201911007178 A CN 201911007178A CN 110852352 B CN110852352 B CN 110852352B
Authority
CN
China
Prior art keywords
target
image
coordinates
max
data
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
Application number
CN201911007178.8A
Other languages
Chinese (zh)
Other versions
CN110852352A (en
Inventor
田大湧
周德云
谢照
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Northwestern Polytechnical University
Original Assignee
Northwestern Polytechnical University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Northwestern Polytechnical University filed Critical Northwestern Polytechnical University
Priority to CN201911007178.8A priority Critical patent/CN110852352B/en
Publication of CN110852352A publication Critical patent/CN110852352A/en
Application granted granted Critical
Publication of CN110852352B publication Critical patent/CN110852352B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Abstract

The invention discloses a data enhancement method for training a target detection deep neural network model, which enhances data by adding three elements of three-dimensional change, lens distortion and illumination change into the traditional data enhancement method, so that a data set contains more pictures, and the enhanced new data is used for training the deep neural network to have higher precision in target detection.

Description

Data enhancement method for training deep neural network model for target detection
Technical Field
The invention belongs to the field of target detection, and particularly relates to a data enhancement method for training a target detection deep neural network model.
Background
Data enhancement is an important means for improving the training effect of the deep neural network, and the essence of the data enhancement is to generate data similar to training data so as to solve the problem of insufficient training data; object detection is the detection of objects of interest from an image or a video segment, such as the detection of pedestrians, vehicles and traffic management signs in the video receiver of an unmanned vehicle; at present, two data enhancement methods for a target detection deep neural network model in the market are provided, namely a traditional enhancement method and a network generation-based method;
The conventional enhancement method is to directly carry out transformations such as rotation, scaling, noise addition and the like on existing data in a training set to generate new data, the transformation process of the data is commonly called affine transformation, besides affine transformation, another common new data generation method is a countercheck generation network, the image enhancement effects of the affine transformation and the countercheck generation network are compared and found aiming at the image classification problem, the affine transformation effect is close to that of the countercheck generation network, but the calculation amount of the affine transformation is much smaller;
however, in the prior art, the widely adopted traditional data enhancement method only adopts two-dimensional change of images/videos to obtain new data, the obtained new data has errors compared with real data, and the target detection precision is not high by using a neural network trained by the method.
Disclosure of Invention
In order to overcome the defects of the prior art, the utility model aims to provide a data enhancement method for training of target detection degree of depth neural network model, through adding three-dimensional change in traditional enhancement method, the camera lens distortion, illumination change three key element, can make the new data that generate be close true data more, utilize the neural network of its training, the target detection precision is higher.
In order to achieve the purpose, the technical scheme of the invention is realized as follows:
a data enhancement method for training a target detection deep neural network model comprises the following specific calculation steps:
the method comprises the following steps: reading images in the training set, and performing normalization processing on read image pixels;
step two: extracting all marked targets from the existing training data in the image read in the step one, namely, manually marking partial data in the training data and intercepting all marked data;
step three: randomly zooming the extracted target, namely generating a random number positioned in an interval [ a, b ] by using a computer as a zooming multiple, zooming the extracted target by utilizing a cubic spline interpolation method, and generating a zoomed target;
step four: performing three-dimensional random rotation on the zoomed target, namely generating three located intervals by using a computer
Figure BDA0002243112990000021
Random number of (theta) x ，θ y And theta z The rotation angles of the target around the three axes X, Y and Z are respectively represented, and if the original coordinates of the point on the target in the three-dimensional space are (X, Y and Z) and the coordinates after rotation around the coordinate axes are (X ', Y' and Z '), then (X, Y, Z) and (X', Y ', Z') satisfy the following relations:
Figure BDA0002243112990000031
Step five: mapping the three-dimensionally rotated target to an XY plane where the image is located, namely assigning a pixel value at the coordinate of the original target (x, y) to a pixel point of a new target round (x ', y');
step six: simulating illumination change by adopting a simplified model, and processing the mapped target to obtain new data added with the illumination change;
step seven: adding lens distortion, wherein the specific distortion process is to set the target coordinates after the fifth processing as (x, y) and the coordinates after the radial distortion as (x ', y'), and then tangentially distortion is carried out on the coordinates (x ', y') to obtain (x ', y') so as to generate a final new target;
step eight: and randomly putting the generated new target into the background of the images in the training set.
Preferably, the reading of the images in the training set and the normalization of the images in the step one specifically include:
s1, reading an image in a training set, wherein the pixel value range of the image is [0,225], the image has three channels of R, Y and B, and the image is regarded as a three-dimensional matrix of W multiplied by H multiplied by 3;
s2, carrying out normalization processing on the image pixel value, namely dividing the image pixel value by 255 to change the value range of the pixel value into [0,1 ];
wherein: w is the image width and H is the image height.
3. The method for enhancing data of target detection deep neural network model training as claimed in claim 1, wherein the specific process of transforming the three-dimensional original coordinates of the points on the target in step four from (x, y, z) around the coordinate axis to (x ', y ', z ') is as follows: and (3) rotating the point (X, Y, Z) on the target around the X axis, then rotating around the Y axis, and finally rotating around the Z axis to obtain a new coordinate point (X ', Y ', Z ').
Preferably, the specific processing procedure of adding the new data obtained by change in the step six is as follows:
s1, randomly generating four numbers mu by a computer 1 ,μ 2 ∈(a μ ,b μ ),σ 1 ,σ 2 ∈(a σ ,b σ ) Respectively as the mean and standard deviation of the two-dimensional normal distribution probability density function, wherein the two-dimensional normal distribution probability density function is as follows:
Figure BDA0002243112990000041
s2, randomly generating four numbers x by a computer min ,x max ∈(0,w'),y min ,y max ∈(0,h')；
Wherein: w 'and h' are the maximum width and the maximum height of the newly generated target matrix;
s3, dividing the interval [ x ] min ,x max ]And [ y min ,y max ]Discretizing into a set { x min ,x min+1 ,x min+2 ,···,x max And { y } min ,y min+1 ,y min+2 ,···,y max Is given by [ x ] min ,x max ]Total M elements, [ y ] min ,y max ]A total of N elements;
s4, sampling two-dimensional normal distribution probability density functions on the two sets to obtain an MXN two-dimensional matrix S; s m th row and n th column of f (x) min+m ,y min+n )，
Wherein: f is a two-dimensional normal distribution probability density function;
S5, the newly generated target is also an image and is provided with three channels of R, G and B, the pixel value of each channel is a matrix, and the dot product of the matrix formed by the pixel values of each channel and the S is respectively calculated, namely the multiplication of the corresponding position elements of the matrix is carried out, so that the new target added with the illumination change can be obtained.
Preferably, the relationship between the target coordinates (x, y) and the coordinates (x ', y') after radial distortion in step seven satisfies:
x'＝x(1+k 1 r 2 +k 2 r 4 +k 3 r 6 )
y'＝y(1+k 1 r 2 +k 2 r 4 +k 3 r 6 )
wherein: r is the distance from a point to the center of the target, k 1 ,k 2 And k 3 Is a preset parameter and can be adjusted according to the equipment for acquiring the image.
Preferably, the relationship between the coordinates (x ', y') of the radial distortion and the coordinates (x ", y") of the tangential distortion satisfies:
x”＝x'+(2p 1 y'+p 2 (r 2 +2x' 2 ))
y”＝y'+(2p 1 x'+p 1 (r 2 +2y' 2 ))
wherein: p is a radical of 1 、p 2 Is a preset parameter and can be adjusted according to the equipment for acquiring the image.
The invention has the beneficial effects that: the invention provides a data enhancement method for training a target detection deep neural network model, which is characterized in that three factors of three-dimensional change, lens distortion and illumination change are added in the traditional enhancement method to enhance data, so that a data set contains more pictures, and the enhanced new data is used for training the deep neural network to have higher precision in target detection.
Detailed Description
In order to make those skilled in the art better understand the technical solution of the present invention, the following further describes the technical solution of the present invention with reference to the embodiments.
A data enhancement method for training a target detection deep neural network model comprises the following specific calculation steps:
the method comprises the following steps: reading images in a training set, and carrying out normalization processing on extracted image pixels, wherein the specific steps are as follows;
s1, reading an image in a training set, wherein the pixel value range of the image is [0.255], the image has three channels of R, G and B, and the image is regarded as a three-dimensional matrix of W multiplied by H multiplied by 3;
wherein: w is the image width, H is the image height;
s2, carrying out normalization processing on the pixel values, namely dividing the pixel values by 255 to change the value range of the pixel values into [0, 1 ].
Step two: extracting all marked targets from the existing training data of the image, namely, manually marking partial data in the training data and intercepting all marked data; the manual marking and the extraction method are the optimal extraction method of the invention, the implementation is simple, and the marking principle is as follows: if the manual marking is that the coordinates of the upper left corner of the area are (100, 200) and the coordinates of the lower right corner are (300, 400), then the rectangular areas of the two points contain the marking target; besides the method, other marking methods can be used, such as marking the target by using a matrix with only 0 and 1 values, wherein the position with 1 represents the target, and the position with 0 represents the background.
Step three: randomly zooming the extracted target, namely generating a random number positioned in an interval [ a, b ] by using a computer as a zooming multiple, zooming the extracted target by utilizing a cubic spline interpolation method, and generating a zoomed target; wherein [ a, b ] is the value range of the random number, which is adjusted according to the characteristics of the data set itself, and is not a fixed value.
Step four: performing three-dimensional random rotation on the zoomed target, namely generating three located intervals by using a computer
Figure BDA0002243112990000061
Random number of (theta) x ，θ y And theta z The rotation angles of the target around the three axes X, Y and Z are respectively represented, and if the original coordinates of the point on the target in the three-dimensional space are (X, Y and Z) and the coordinates after rotation around the coordinate axes are (X ', Y' and Z '), then (X, Y, Z) and (X', Y ', Z') satisfy the following relations:
Figure BDA0002243112990000062
step five: and mapping the three-dimensionally rotated target to an XY plane where the image is located, namely, assigning a pixel value at the coordinate of the original target (x, y) to a pixel point of the new target round (x ', y'), wherein for example, the pixel value at the original target (142,173) is (0.5,0.1,0.4), the pixel value at the new target (100.200) is (0.5,0.1,0.4) after the three-dimensional rotation and the mapping to the XY plane, and the point (142,173) is transformed to (100.37,200.5).
Wherein: round means round to round;
step six: simulating illumination change by using a simplified model, processing the mapped target to obtain new data added with the illumination change, and specifically comprising the following steps of:
s1, randomly generating four numbers mu by a computer 1 ,μ 2 ∈(a μ ,b μ ),σ 1 ,σ 2 ∈(a σ ,b σ ) Respectively as the mean and standard deviation of the two-dimensional normal distribution probability density function, wherein the two-dimensional normal distribution probability density function is as follows:
Figure BDA0002243112990000071
s2, randomly generating four numbers x by a computer min ,x max ∈(0,w'),y min ,y max ∈(0,h')；
Wherein: w 'and h' are the maximum width and the maximum height of the newly generated target matrix;
s3, dividing the interval [ x ] min ,x max ]And [ y min ,y max ]Discretizing into a set { x min ,x min+1 ,x min+2 ,···,x max And { y } min ,y min+1 ,y min+2 ,···,y max Is given by [ x ] min ,x max ]Total M elements, [ y ] min ,y max ]A total of N elements;
s4, sampling two-dimensional normal distribution probability density functions on the two sets to obtain an MXN two-dimensional matrix S; s m th row and n th column of f (x) min+m ,y min+n )，
Wherein: f is a two-dimensional normal distribution probability density function;
s5, a newly generated target is also an image and is provided with three channels of R, G and B, the pixel value of each channel is a matrix, the dot product of the matrix formed by the pixel values of each channel and S is respectively calculated, namely the pixel value of the corresponding position of the matrix is multiplied, so that a new target matrix S ' w ' h ' with illumination change can be obtained, because the target is essentially an area in the image, the target is also provided with three R, G, B channels, the target can be represented into a matrix with width x height 3 no matter whether the target is a new target or an untransformed target, w ' is the width of the new target, and h ' is the height of the new target;
The S-M-N is a sample of a normal distribution function, and since the normal distribution function is continuous, only a discrete interval is available, for example, a function in the interval of (-1,1) is sampled into 20 points, and then function values at 20 points of-1, -0.9, -0.8,. 9, 0.8, 0.9, 1 are retained; the same continuous interval can sample different numbers of function values, such as-1, -0.8, -0.6, …,0.6,0.8,1, so that only 10 points are sampled;
carrying out different sampling on the continuous two-dimensional normal distribution function to generate matrixes with different sizes; the generated matrix is dot-product with each channel of the new target, i.e. the two matrices are multiplied by corresponding position elements, so that the matrix obtained by sampling the normal distribution should be the same as the maximum width and the maximum height of the new target.
Step seven: adding lens distortion:
s1, radial distortion, namely distortion distributed along the boundary direction of the lens, wherein the target coordinates after the first step to the fifth step are (x, y), and the coordinates after the radial distortion are (x ', y'), which satisfy the following relations:
x'＝x(1+k 1 r 2 +k 2 r 4 +k 3 r 6 )
y'＝y(1+k 1 r 2 +k 2 r 4 +k 3 r 6 )
wherein: r is the distance from a point to the center of the target, k 1 ,k 2 And k 3 The parameters are preset parameters and can be adjusted according to equipment for acquiring images;
S2, tangential distortion, namely, setting the coordinates subjected to radial distortion as (x ', y '), and the coordinates subjected to tangential distortion as (x ', y "), wherein the coordinates satisfy the following relation:
x”＝x'+(2p 1 y'+p 2 (r 2 +2x' 2 ))
y”＝y'+(2p 1 x'+p 1 (r 2 +2y' 2 ))
wherein: p is a radical of 1 、p 2 Is a preset parameter and can be adjusted according to the equipment for acquiring the image.
Finally, the pixel value of the target coordinate (x, y) mapped in the step five is assigned to (x', y ") to generate a final new target;
step eight: and randomly putting the generated new target into the background of the images in the training set.
The first embodiment is as follows: in practice, due to the fact that the data volumes of data sets of different classes are different, a data enhancement technology (data augmentation technology) is used in training, we simply ignore the classes with the instances smaller than 100, the remaining 45 classes can be used for classification, the classes with the instances between 100 and 1000 are enhanced to 1000 by the method, and the rest is more than 1000;
then, a traffic sign detection algorithm based on a deep neural network is trained by using a newly generated data set, the training effect is evaluated by using the average accuracy, and the accuracy of the traffic sign detection algorithm is compared with the accuracy of a neural network trained by using a traditional data enhancement method, as shown in table 1:
TABLE 1 average accuracy comparison (%)
Figure BDA0002243112990000091
Figure BDA0002243112990000101
As can be seen from the table I, the average accuracy of the training data enhanced by the method is generally higher than that of the neural network trained by the traditional data enhancement method.
The foregoing shows and describes the general principles, essential features, and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims (5)

1. A data enhancement method for training a target detection deep neural network model is characterized by comprising the following specific calculation steps:
the method comprises the following steps: reading images in the training set, and performing normalization processing on pixels of the read images;
step two: extracting all marked targets from the existing training data in the image read in the step one, namely, manually marking partial data in the training data and intercepting all marked data;
step three: randomly zooming the extracted target, namely generating a random number positioned in an interval [ a, b ] by using a computer as a zooming multiple, zooming the extracted target by utilizing a cubic spline interpolation method, and generating a zoomed target;
Step four: performing three-dimensional random rotation on the zoomed target, i.e. by using computerGenerate three location intervals
Figure FDA0003573731560000011
Random number of (theta) x ，θ y And theta z Respectively representing the rotation angles of the target around three axes of X, Y and Z, and setting the original coordinate of a point on the target in a three-dimensional space as (X) o ，y o Z) with coordinates after rotation about the coordinate axis being (x, y, z'), then (x) o ，y o Z) and (x, y, z') satisfy the following relationship:
Figure FDA0003573731560000012
step five: mapping the target after three-dimensional random rotation to an XY plane where the image is positioned, namely assigning the pixel value of the original target (x, y) coordinate to a new target round (x) o ，y o ) The pixel point of (2);
step six: simulating illumination change by using a simplified model, and processing the mapped target to obtain new data added with the illumination change, wherein the specific processing process is as follows:
s1, randomly generating four numbers mu by a computer 1 ,μ 2 ∈(a μ ,b μ ),σ 1 ,σ 2 ∈(a σ ,b σ ) Respectively as the mean and standard deviation of the two-dimensional normal distribution probability density function, wherein the two-dimensional normal distribution probability density function is as follows:
Figure FDA0003573731560000021
s2, randomly generating four numbers x by a computer min ,x max ∈(0,w'),y min ,y max ∈(0,h')；
Wherein: w 'and h' are the maximum width and the maximum height of the newly generated target matrix respectively;
s3, dividing the interval [ x ] min ,x max ]And [ y min ,y max ]Discretizing it intoSet { x min ,x min+1 ,x min+2 ,···,x max And { y } min ,y min+1 ,y min+2 ,···,y max Is given by [ x ] min ,x max ]Total M elements, [ y ] min ,y max ]A total of N elements;
s4, sampling two-dimensional normal distribution probability density functions on the two sets to obtain an MXN two-dimensional matrix S; s m th row and n th column of f (x) min+m ,y min+n )，
Wherein: f is a two-dimensional normal distribution probability density function;
s5, the newly generated target is also an image and is provided with three channels of R, G and B, the pixel value of each channel is a matrix, and the dot product of the matrix formed by the pixel values of each channel and the S is respectively calculated, namely the multiplication of the corresponding position elements of the matrix is carried out, so that the new target with the illumination change can be obtained;
step seven: adding lens distortion, wherein the specific distortion process is to set the target coordinates after the fifth processing as (x, y) and the coordinates after the radial distortion as (x ', y'), and then tangentially distortion is carried out on the coordinates (x ', y') to obtain (x ', y') so as to generate a final new target;
step eight: and randomly placing the generated new target in the background of the images in the training set.
2. The method as claimed in claim 1, wherein the step one of reading the images in the training set and normalizing the pixels of the read images comprises the specific steps of:
S1, reading an image in a training set, wherein the pixel value range of the image is [0,225], the image has three channels of R, G and B, and the image is regarded as a three-dimensional matrix of W multiplied by H multiplied by 3;
s2, carrying out normalization processing on the image pixel value, namely dividing the image pixel value by 255 to change the value range of the pixel value into [0,1 ];
wherein: w is the image width and H is the image height.
3. The method of claim 1, wherein the three-dimensional original coordinates of the points on the target in step four are represented by (x) o ，y o Z) the specific process of rotation around coordinate axis transformation to (x, y, z') is: point (x) on the target o ，y o And Z) rotating around the X axis, then rotating around the Y axis, and finally rotating around the Z axis to obtain a new coordinate point (X, Y, Z').
4. The method as claimed in claim 1, wherein the relationship between the target coordinates (x, y) and the radially distorted coordinates (x ', y') in step seven satisfies the following relationship:
x'＝x(1+k 1 r 2 +k 2 r 4 +k 3 r 6 )
y'＝y(1+k 1 r 2 +k 2 r 4 +k 3 r 6 )
wherein: r is the distance from a point to the center of the target, k 1 ,k 2 And k 3 Is a preset parameter and can be adjusted according to the equipment for acquiring the image.
5. The method of claim 4, wherein the relationship between the coordinates (x ', y') after the radial distortion and the coordinates (x ", y") of the tangential distortion satisfies the following relationship:
x”＝x'+(2p 1 y'+p 2 (r 2 +2x' 2 ))
y”＝y'+(2p 1 x'+p 1 (r 2 +2y' 2 ))
Wherein: p is a radical of 1 、p 2 Is a preset parameter and can be adjusted according to the equipment for acquiring the image.
CN201911007178.8A 2019-10-22 2019-10-22 Data enhancement method for training deep neural network model for target detection Active CN110852352B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911007178.8A CN110852352B (en) 2019-10-22 2019-10-22 Data enhancement method for training deep neural network model for target detection

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911007178.8A CN110852352B (en) 2019-10-22 2019-10-22 Data enhancement method for training deep neural network model for target detection

Publications (2)

Publication Number Publication Date
CN110852352A CN110852352A (en) 2020-02-28
CN110852352B true CN110852352B (en) 2022-07-29

Family

ID=69597191

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911007178.8A Active CN110852352B (en) 2019-10-22 2019-10-22 Data enhancement method for training deep neural network model for target detection

Country Status (1)

Country Link
CN (1) CN110852352B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112183500A (en) * 2020-11-27 2021-01-05 北京易真学思教育科技有限公司 Data enhancement method, training method, computer device and storage medium

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108230233A (en) * 2017-05-16 2018-06-29 北京市商汤科技开发有限公司 Data enhancing, treating method and apparatus, electronic equipment and computer storage media
CN109190768A (en) * 2018-08-09 2019-01-11 北京中关村科金技术有限公司 A kind of data enhancing corpus training method in neural network
JP2019032821A (en) * 2017-06-26 2019-02-28 コニカ ミノルタ ラボラトリー ユー．エス．エー．，インコーポレイテッド Data augmentation techniques using style transformation with neural network
CN109543684A (en) * 2018-10-09 2019-03-29 广州大学 Immediate targets tracking detection method and system based on full convolutional neural networks
CN109801365A (en) * 2019-01-25 2019-05-24 上海小萌科技有限公司 A kind of three-dimensional modeling data acquisition device and its three-dimensional modeling method
CN110175982A (en) * 2019-04-16 2019-08-27 浙江大学城市学院 A kind of defect inspection method based on target detection
WO2019169895A1 (en) * 2018-03-09 2019-09-12 华南理工大学 Fast side-face interference resistant face detection method
CN110348376A (en) * 2019-07-09 2019-10-18 华南理工大学 A kind of pedestrian's real-time detection method neural network based

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110298897A1 (en) * 2010-06-08 2011-12-08 Iva Sareen System and method for 3d virtual try-on of apparel on an avatar

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108230233A (en) * 2017-05-16 2018-06-29 北京市商汤科技开发有限公司 Data enhancing, treating method and apparatus, electronic equipment and computer storage media
JP2019032821A (en) * 2017-06-26 2019-02-28 コニカ ミノルタ ラボラトリー ユー．エス．エー．，インコーポレイテッド Data augmentation techniques using style transformation with neural network
WO2019169895A1 (en) * 2018-03-09 2019-09-12 华南理工大学 Fast side-face interference resistant face detection method
CN109190768A (en) * 2018-08-09 2019-01-11 北京中关村科金技术有限公司 A kind of data enhancing corpus training method in neural network
CN109543684A (en) * 2018-10-09 2019-03-29 广州大学 Immediate targets tracking detection method and system based on full convolutional neural networks
CN109801365A (en) * 2019-01-25 2019-05-24 上海小萌科技有限公司 A kind of three-dimensional modeling data acquisition device and its three-dimensional modeling method
CN110175982A (en) * 2019-04-16 2019-08-27 浙江大学城市学院 A kind of defect inspection method based on target detection
CN110348376A (en) * 2019-07-09 2019-10-18 华南理工大学 A kind of pedestrian's real-time detection method neural network based

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
3D Steerable CNNs: Learning Rotationally Equivariant Features in Volumetric Data;Maurice Weiler 等;《32nd Conference on Neural Information Processing Systems》;20181231;第1-12页 *
A Data Augmentation Strategy Based on Simulated Samples for Ship Detection in RGB Remote Sensing Images;Yiming Yan 等;《International Journal of Geo-Information》;20190613;第1-18页 *
Learning Data Augmentation Strategies for Object Detection;Barret Zoph 等;《arXiv:1906.11172v1》;20190630;第1-13页 *
基于Faster R-CNN及数据增广的满文文档印章检测;卢海涛 等;《大连民族大学学报》;20180930;第20卷(第5期);第455-459页 *
基于Faster-RCNN的遥感图像飞机检测算法;张中宝 等;《南京师大学报(自然科学版)》;20181231;第41卷(第4期);第79-86页 *

Also Published As

Publication number Publication date
CN110852352A (en) 2020-02-28

Similar Documents

Publication Publication Date Title
CN110264416B (en) Sparse point cloud segmentation method and device
CN109886896B (en) Blue license plate segmentation and correction method
CN109272455B (en) Image defogging method based on weak supervision generation countermeasure network
CN111223053A (en) Data enhancement method based on depth image
CN106408597A (en) Neighborhood entropy and consistency detection-based SAR (synthetic aperture radar) image registration method
CN110807807B (en) Monocular vision target positioning pattern, method, device and equipment
CN111369605B (en) Infrared and visible light image registration method and system based on edge features
CN109919832B (en) Traffic image splicing method for unmanned driving
CN106156684A (en) A kind of two-dimensional code identification method and device
CN113034593B (en) 6D pose labeling method, system and storage medium
CN108416801B (en) Har-SURF-RAN characteristic point matching method for stereoscopic vision three-dimensional reconstruction
CN111382658B (en) Road traffic sign detection method in natural environment based on image gray gradient consistency
CN104008542A (en) Fast angle point matching method for specific plane figure
CN111739031A (en) Crop canopy segmentation method based on depth information
CN113592923A (en) Batch image registration method based on depth local feature matching
CN110852352B (en) Data enhancement method for training deep neural network model for target detection
CN107463939B (en) Image key straight line detection method
CN109635809B (en) Super-pixel segmentation method for visual degradation image
CN111797832B (en) Automatic generation method and system for image region of interest and image processing method
CN112686247A (en) Identification card number detection method and device, readable storage medium and terminal
CN110070626B (en) Three-dimensional object retrieval method based on multi-view classification
CN116206156A (en) Pavement crack classification and identification method under shadow interference
CN112396638A (en) Image processing method, terminal and computer readable storage medium
CN104408430B (en) License plate positioning method and device
Li et al. Feature point extraction and tracking based on a local adaptive threshold

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