CN108460426A - A kind of image classification method based on histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder - Google Patents
A kind of image classification method based on histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder Download PDFInfo
- Publication number
- CN108460426A CN108460426A CN201810269829.XA CN201810269829A CN108460426A CN 108460426 A CN108460426 A CN 108460426A CN 201810269829 A CN201810269829 A CN 201810269829A CN 108460426 A CN108460426 A CN 108460426A
- Authority
- CN
- China
- Prior art keywords
- pseudoinverse
- learning
- encoding encoder
- training
- image
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention discloses a kind of image classification method based on histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder, image gradient features are extracted using histograms of oriented gradients (HOG) including (1), calculate the directional diagram of image, the direction character that several overlapping regional areas are counted by HOG operators, obtains the HOG features of image.The parameter that we set different HOG operators obtains several HOG features, by these Fusion Features at the feature vector of higher-dimension.(2) pseudoinverse learning algorithm training storehouse self-encoding encoder (PILAE) is used, the high dimensional feature being fused into upper step is put into PILAE and continues learning characteristic.(3) feature learnt in PILAE is put into grader and is classified.HOG can extract the information of two-dimensional image.Pseudoinverse learning algorithm is a kind of non-iterative method, is used for Training Multilayer Feedforward Neural Networks.Model proposed by the present invention, training time are advantageous compared with other models, and most of hyper parameters are decided in its sole discretion by input data and network structure, need not be manually set.
Description
Technical field
The invention belongs to field of artificial intelligence, are related to a kind of quick training depth nerve net of image characteristics extraction combination
The method of network, the more particularly to model of histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder are used for image classification
Task.
Background technology
In recent years, the development of deep learning causes the upsurge of artificial intelligence study again.Deep learning originates from previous generation
Artificial intelligence technology --- the artificial nerve network model recorded the forties.First neural network model perceptron algorithm is by instruction
White silk, which can reach, classifies to certain input vector model, but later it has been proposed that the function of simple perceptron algorithm is
It is limited, linearly inseparable problem can not be handled, artificial neural network research is initially located in low tide one.In recent years, due to calculating
Machine hardware advances are rapid, and the calculated level of computer is greatly improved so that deep learning is risen again.
Deep learning handles data flow by simulating human brain, is that use includes labyrinth from the training data provided
Or the method that multiple processing layer datas carry out higher level of abstraction is constituted by multiple nonlinear transformation.Currently, deep learning is regarded in image
Frequently it, made breakthrough progress in sound and text identification, there is stronger learning ability than shallow-layer learning algorithm.In addition to strong
Big Function Fitting ability and good generalization ability, every layer of learning table of deep learning have revealed from low-level image feature to high-level characteristic
Process.
In the Internet era of data explosion, the largely data without mark are generated daily, and train depth model often
The data for needing largely to carry label, create a complete tape label data set need to expend a large amount of manpower and when
Between.Currently, deep learning Most models are all to have supervision, such as the multilayer feedforward neural network of tape label, convolutional Neural net
Network etc.;Unsupervised model is mainly self-encoding encoder.Unsupervised learning is advantageous in that data need not mark, and model can be automatic
Learning characteristic therefore develop the future thrust that unsupervised deep learning is deep learning.
Invention content
(1) in view of this, the technical problem to be solved in the present invention is to provide one kind combining puppet based on histograms of oriented gradients
The model of storehouse self-encoding encoder composition is trained in reversal learning, is used for image classification.The model is extracted using image manual features first
Method histograms of oriented gradients extracts feature, and the storehouse self-encoding encoder for connecting pseudoinverse learning training later further extracts feature,
Then feature is put into grader and is classified.
(2) present invention is to solve the above problems, the following technical solutions are proposed:
1) Gradient Features of histograms of oriented gradients (HOG) extraction image are used.Histograms of oriented gradients is that computer regards
Operator is described for image detection clarification of objective in feel.HOG obtains whole picture by the Gradient direction information of statistical picture part
The shape and profile information feature of image.HOG divides an image into several networks (block), cell factory first
(cell) gradient information is calculated in each block, in order to preferably describe image gradient features, HOG uses local overlapping
It calculates.Since the feature of different cell sizes, different block sizes and different sliding steps extraction is different.We make
Different HOG features are extracted with several different size of cell, block and sliding step parameter, then spell these features
It is connected into the feature vector of a higher-dimension.This thought is similar in convolutional neural networks, is extracted using different convolution kernels different
Feature map.
2) use pseudoinverse learning algorithm (Pseudoinverse Learning Algorithm, PIL) training storehouse self-editing
Code device.The high dimensional feature vector that upper step obtains is put into self-encoding encoder by we continues learning characteristic.Self-encoding encoder using
The input of unsupervised learning mode, the network is approximately equal to output, and the quantity by controlling hidden node can make input vector rise dimension
(increasing the number of hidden nodes) or dimensionality reduction (reducing hidden layer energy number of nodes), we reduce hidden node quantity here so that self-editing
The code automatic learning characteristic of device, certainly, we are added to weight attenuation term in loss function, avoid the identical study of self-encoding encoder.
General trained self-encoding encoder needs to set loss function, and optimization is iterated using gradient descent algorithm.In order to save training
Time, we are using pseudoinverse learning algorithm training storehouse self-encoding encoder here.Pseudoinverse learning algorithm is in nineteen ninety-five, by Guo Pingjiao
Award a kind of algorithm (Guo et al, " An Exact for efficient training Single hidden layer feedforward neural networks of proposition
Supervised Learning for a Three-Layer Supervised Neural Network”,ICONIP'95,
Pp.1041-1044,1995.), multilayer neural network (Guo et al, " Pseudoinverse were expanded in 2001
Learning Algorithm for Feedforward Neural Networks”,in Mastorakis Eds.,
Advances in Neural Networks and Applications,WSES Press(Athens)pp.321-326,
2001.).PIL algorithm ideas are to replace weight using the pseudo inverse matrix of input matrix, substitute gradient descent algorithm.PIL passes through square
Battle array operation, quickly trains multilayer neural network, more efficient compared to gradient descent algorithm.Specifically, when the nerve of a multilayer
Network has N number of sample, each sample to have m dimensions, and input matrix X, expectation target O, we train the purpose of neural network
It is to find one group of parameter so that the value of loss function reaches minimum:
Wherein, g (x, Θ) is the mapping function of neural network, and Θ is parameter sets.We can be represented by the following formula more
Relationship between l layers and l+1 layers of neural network of layer.
Yl+1=σ (YlWl),
Wherein σ () is activation primitive, YlRepresent l layers of output.In this way last layer network export us can be as
Lower expression:
G=YLWL,
G is target output, W in this formulaLRepresent the weight of last layer.In summary three formulas, Wo Menke
Loss function is rewritten as following formula:
In this way, this problem becomes Linear least squares minimization problem.The optimal pseudoinverse solution of above formula is W=(YL)+O, optimal
Solution brings loss function into, and loss function can be written as form again by us:
From the above equation, we can see that only needing YL(YL)+Close to unit matrix I, loss function is optimal solution, and therefore, we reset
Optimization aimWherein e is the error threshold that we set, as long as square error is less than this in the calculation
Threshold value is considered as and reaches optimization aim.In training process, every layer is calculated | | Yl(Yl)+-I||2, if it is less than the threshold value of setting
E, training are completed, and algorithm terminates, and otherwise increase a hidden layer, continue to calculate square error, the threshold value model until converging to setting
In enclosing or the hidden layer number of training reaches the number of plies of setting.
In the training process, regularization term is added in we in loss function, avoids causing to intend due to data volume deficiency
It closes.After introducing weight decaying regularization, optimization aim is changed to:
Wherein, λ > 0 are regularization coefficient, by above formula, it can be deduced that (YL)+=(YYT+λΙ)-1YT。
(3) the advantage of the invention is that:
Hyper parameter need not be set:The number of HOG operators, cell sizes and block sizes by setting by hand in the present invention;
Hidden layer number is determined by loss function in PILAE, the training stopping when error rises or reaches the number of plies set by user;Hidden unit
Number is determined that the intrinsic dimensionality for commonly entering matrix is more than rank of matrix, and the dimension that we reduce matrix disappears by the order of input matrix
Except redundant data, correlated characteristic is reduced, when the intrinsic dimensionality of input matrix is equal to rank of matrix, we are with certain proportion
The dimension for forcing reduction matrix, achievees the purpose that feature learning;Since PILAE does not need gradient optimization algorithm and iteration is excellent
Change, so there is no need to set learning rate and study rounds;The weight of neural network of the present invention is the pseudo inverse matrix of input matrix, therefore not
It needs to initialize weight;
Training time is short:The present invention using the different HOG features of different HOG operator extractions finally form high dimensional feature to
Amount.Given HOG operator parameters only need the feature that picture extracts merging features into higher-dimension one by one.Connection PILAE is carried out below
The training of further feature learning, wherein PILAE does not need iteration optimization, only needs linear algebra calculating that training can be completed, compares
Other Model of Neural Network iterate, undated parameter, and the present invention has advantage on the training time.And in model of the present invention not
It needs to adjust hyper parameter, so for the user, not only the training time is fast, but also saves many and spend in the super ginseng of debugging
Several time.
AI democratizes:Nowadays, many network models achieve the error rate of very little on image classification data collection, but work as
Network model is applied in the actual use scene of oneself by user, often cannot get ideal effect.It often also needs to adjust
Many network parameters, training is for a long time.It is a thing for being difficult for professional to adjust neural network parameter, more
Let alone the people of not specialty background.The present invention is more conducive to AI democratizations using complicated tune ginseng simply, is not needed.
Description of the drawings
Fig. 1 histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder structural schematic diagrams
Specific implementation mode
(1) below with reference to attached drawing, the preferred embodiment of the present invention is described in detail:
The present invention proposes a kind of image based on histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder
Sorting technique.In order to make the purpose , technical scheme and advantage of the present invention be clearer, below in conjunction with specific implementation example and
This method is described in further detail in attached drawing.It should be appreciated that herein specific implementation example description only to explain the present invention, and
It is not used in the restriction present invention.
Specifically, it is that one kind of embodiment of the present invention being based on histograms of oriented gradients pseudo- reversal learning is combined to instruct shown in Fig. 1
Practice storehouse self-encoding encoder and combines the method for being used for Handwritten Digit Classification.For the picture group of N given n × n-pixel sizes
At training sample set X, matrix is expressed as X=[x1,x2,…,xN].The present invention embodiment based on histograms of oriented gradients
Include following basic step in conjunction with pseudoinverse learning training storehouse self-encoding encoder image classification method:
Step 1) selects t HOG to describe operator, one by one image zooming-out HOG features, the parameter cell sizes of these operators,
Block sizes and gradient direction are different.The feature vector that the feature vector of t operator extraction is connected into m dimensions forms spy
Sign matrix F is for training PILAE, matrix to be expressed as F=[f1,f2,…,fN]。
Step 2) solves the pseudo inverse matrix F of F using features described above matrix F as the input matrix of self-encoding encoder+, first by F
Singular value decomposition is carried out to obtain
F=U Σ VT
The order r=Rank (Σ) of input matrix, wherein Rank () function be calculate Σ in be 0 element number, input
The dimension m=Dim (f) of feature vector, wherein Dim () function are to calculate feature vector Characteristic Number.Self-encoding encoder is arranged in we
Hidden unit number p be r<p<m.If rank of matrix r is less than the dimension m of feature vector, Hidden unit number p is set in
r<p<Between m:
P=r+ α (m-r)
Wherein, α is the parameter that user makes by oneself.When matrix full rank, i.e. rank of matrix and intrinsic dimensionality is equal, for characterology
It practises, forcing, which reduces intrinsic dimensionality, makes p<m:
P=β m
Wherein, β is the parameter that user makes by oneself.
F is carried out SVD decomposition first, obtained by step 3) according to PIL algorithms
By F=U Σ VT Wherein, Σ ' is the inverse for the element being not zero in Σ.
Matrix after being blocked for V,
V=[v1,v2..., vp..., vm]T,
Wherein p is the Hidden unit number being arranged in step 2).Then, W is enablede=F+, matrix is mapped to the feature sky of hidden layer
Between in:
H=σ (WeF)
Wherein σ () is activation primitive.
Step 4) solves decoder weight according to pseudo- reversal learning.Self-encoding encoder decoder weight WdH=X, according to minimum two
Multiply that there are optimal pseudoinverse approximate solution Wd=XH+, therefore calculate the pseudoinverse H of hidden layer output H+.The loss function definition of pseudo- reversal learning is such as
Under:
MinE=| | X-WH | |2
In order to avoid model over-fitting, we increase weight decaying regularization term, and loss function is amended as follows:
Loss function minimum value is solved, is obtained:
MinE=- (X-WH) HT+ kW=0
W=XHT(ΗΗT+kI)-1
Step 5) obtains decoder weight W by step 4)d, using the transposition of decoder weight as the weight of encoderIn this way, the output H=σ (W of the hidden layer of self-encoding encodereF the character representation for) indicating initial data, hidden layer is exported
As the input data of next self-encoding encoder, the next self-encoding encoder of step (1-4) training is repeated.When the requirement for reaching user
Afterwards, deconditioning opens trained self-encoding encoder, forms storehouse self-encoding encoder, decoder section is removed, last is defeated
It is exactly the feature of initial data to go out, and is put into grader classifies later.
(2) embodiment
In order to prove that the present invention is practical, we use the performance of the common data set testing model of machine learning.And
Experiment is compared with correlation model.
Experiment database used is handwritten form database (THE MNIST DATABASE of handwritten
Digits), the existing standard data sets that detection sorting algorithm function admirable is known as by industry of MNIST.We are detected using MNIST
The model performance of the present invention.MNIST by Yann LeCun et al. create include 0~9 handwriting digital image data set, data
Collection includes altogether 70000 handwriting digital images, wherein 60000 training images, 10000 detection images, every image is all
It is pre-processed through past background, and by its image to 28 × 28=784 pixel.We use classical engineering
Practise, the model of neural network model and the present invention be based on histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder into
Row comparison, as a result as illustrated in chart 1.
Comparing remaining model and can be seen that model proposed by the present invention has apparent advantage on the training time, and takes
Obtained good accuracy of identification.
Model | Time consumption for training (second) | Training precision (%) | Measuring accuracy (%) |
SAE | 298.43 | 97.53 | 96.72 |
Lenet-5 | 523.43 | 100.00 | 98.33 |
SVM | 2583.82 | 98.72 | 96.46 |
HOG | 30.83 | 94.88 | 94.32 |
PILAE | 62.32 | 97.32 | 96.39 |
HOG+PILAE | 92.58 | 98.82 | 98.01 |
Table 1
The foregoing is merely the preferred embodiment of the present invention, are not intended to restrict the invention, it is clear that those skilled in the art
Various changes and modifications can be made to the invention by member without departing from the spirit and scope of the present invention.If in this way, the present invention
Within the scope of the claims of the present invention and its equivalent technology, then the present invention is also intended to include these these modifications and variations
Including modification and variation.
Claims (5)
1. a kind of image classification method based on histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder, including with
Lower step:
1) Gradient Features of use direction histogram of gradients extraction image, used here as the HOG operator extractions of several different parameters
Then these different features are cascaded into high dimensional feature by different Gradient Features.
2) pseudoinverse learning algorithm training storehouse self-encoding encoder (PILAE) is used, using the high dimensional feature of step 1) as mode input,
Further learning characteristic.
3) feature that PILAE is trained is input in grader and carries out image classification.
2. according to claim 1, use direction histogram of gradients extracts image gradient information, which is characterized in that step
1) histograms of oriented gradients used by is the Gradient Features for extracting image, is calculated its main feature is that setting multiple HOG descriptions
The parameter of son, extracts different Gradient Features, by these Fusion Features at the feature vector of higher-dimension.
3. the image according to claim 1, based on histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder
Sorting technique, which is characterized in that the pseudoinverse learning algorithm employed in step 2) is a kind of no iterative algorithm, for training multilayer
Feedforward neural network is not necessarily to iteration optimization its main feature is that without backpropagation, the weight of network by input matrix pseudo inverse matrix
It determines, using mean square error as loss function, last layer of weight is solved using least square in last layer of network.
4. it is by input matrix according to claim 1, to train the Hidden unit number of self-encoding encoder using pseudoinverse learning algorithm
Order determine that the dimension m=Dim (f) of input feature value, Dim () function is to calculate feature vector Characteristic Number and input
Rank of matrix r=Rank (Σ), wherein Rank () function be calculate in Σ be not 0 element number, m and r are established connection by us
System, sets the Hidden unit number p of self-encoding encoder as r<p<m.If rank of matrix r is less than the dimension m of feature vector, will be hidden
Layer unit number p is set in r<p<Between m:P=r+ α (m-r), wherein α is the parameter that user makes by oneself, when matrix full rank, i.e. matrix
Sum of ranks intrinsic dimensionality it is equal, for feature learning, we, which force to reduce intrinsic dimensionality, makes p<m:P=β m, wherein β is to use
The parameter that family is made by oneself.
5. according to claim 1, the hidden layer number of pseudoinverse learning algorithm training storehouse self-encoding encoder be by the formula that defines from
Dynamic confirmation:Wherein e is the threshold value of setting, calculates every layer of training error, when less than given threshold, training
Stop.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810269829.XA CN108460426A (en) | 2018-03-29 | 2018-03-29 | A kind of image classification method based on histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810269829.XA CN108460426A (en) | 2018-03-29 | 2018-03-29 | A kind of image classification method based on histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108460426A true CN108460426A (en) | 2018-08-28 |
Family
ID=63237188
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810269829.XA Pending CN108460426A (en) | 2018-03-29 | 2018-03-29 | A kind of image classification method based on histograms of oriented gradients combination pseudoinverse learning training storehouse self-encoding encoder |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108460426A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110211205A (en) * | 2019-06-14 | 2019-09-06 | 腾讯科技(深圳)有限公司 | Image processing method, device, equipment and storage medium |
CN115908311A (en) * | 2022-11-16 | 2023-04-04 | 湖北华鑫光电有限公司 | Lens forming detection equipment based on machine vision and method thereof |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106127804A (en) * | 2016-06-17 | 2016-11-16 | 淮阴工学院 | The method for tracking target of RGB D data cross-module formula feature learning based on sparse depth denoising own coding device |
CN106874879A (en) * | 2017-02-21 | 2017-06-20 | 华南师范大学 | Handwritten Digit Recognition method based on multiple features fusion and deep learning network extraction |
US20170328194A1 (en) * | 2016-04-25 | 2017-11-16 | University Of Southern California | Autoencoder-derived features as inputs to classification algorithms for predicting failures |
CN107480777A (en) * | 2017-08-28 | 2017-12-15 | 北京师范大学 | Sparse self-encoding encoder Fast Training method based on pseudo- reversal learning |
CN107609637A (en) * | 2017-09-27 | 2018-01-19 | 北京师范大学 | A kind of combination data represent the method with the raising pattern-recognition precision of pseudo- reversal learning self-encoding encoder |
-
2018
- 2018-03-29 CN CN201810269829.XA patent/CN108460426A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170328194A1 (en) * | 2016-04-25 | 2017-11-16 | University Of Southern California | Autoencoder-derived features as inputs to classification algorithms for predicting failures |
CN106127804A (en) * | 2016-06-17 | 2016-11-16 | 淮阴工学院 | The method for tracking target of RGB D data cross-module formula feature learning based on sparse depth denoising own coding device |
CN106874879A (en) * | 2017-02-21 | 2017-06-20 | 华南师范大学 | Handwritten Digit Recognition method based on multiple features fusion and deep learning network extraction |
CN107480777A (en) * | 2017-08-28 | 2017-12-15 | 北京师范大学 | Sparse self-encoding encoder Fast Training method based on pseudo- reversal learning |
CN107609637A (en) * | 2017-09-27 | 2018-01-19 | 北京师范大学 | A kind of combination data represent the method with the raising pattern-recognition precision of pseudo- reversal learning self-encoding encoder |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110211205A (en) * | 2019-06-14 | 2019-09-06 | 腾讯科技(深圳)有限公司 | Image processing method, device, equipment and storage medium |
WO2020248898A1 (en) * | 2019-06-14 | 2020-12-17 | 腾讯科技(深圳)有限公司 | Image processing method, apparatus and device, and storage medium |
CN110211205B (en) * | 2019-06-14 | 2022-12-13 | 腾讯科技(深圳)有限公司 | Image processing method, device, equipment and storage medium |
US11663819B2 (en) | 2019-06-14 | 2023-05-30 | Tencent Technology (Shenzhen) Company Limited | Image processing method, apparatus, and device, and storage medium |
CN115908311A (en) * | 2022-11-16 | 2023-04-04 | 湖北华鑫光电有限公司 | Lens forming detection equipment based on machine vision and method thereof |
CN115908311B (en) * | 2022-11-16 | 2023-10-20 | 湖北华鑫光电有限公司 | Lens forming detection equipment and method based on machine vision |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Mascarenhas et al. | A comparison between VGG16, VGG19 and ResNet50 architecture frameworks for Image Classification | |
CN108717568B (en) | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network | |
CN105678284B (en) | A kind of fixed bit human body behavior analysis method | |
CN105184298B (en) | A kind of image classification method of quick local restriction low-rank coding | |
CN111753828B (en) | Natural scene horizontal character detection method based on deep convolutional neural network | |
CN106295507B (en) | A kind of gender identification method based on integrated convolutional neural networks | |
CN108388896A (en) | A kind of licence plate recognition method based on dynamic time sequence convolutional neural networks | |
CN106126581A (en) | Cartographical sketching image search method based on degree of depth study | |
CN109063719B (en) | Image classification method combining structure similarity and class information | |
CN113627472B (en) | Intelligent garden leaf feeding pest identification method based on layered deep learning model | |
Tang et al. | Deep fishernet for object classification | |
CN109785344A (en) | The remote sensing image segmentation method of binary channel residual error network based on feature recalibration | |
CN110334715A (en) | A kind of SAR target identification method paying attention to network based on residual error | |
CN107767416A (en) | The recognition methods of pedestrian's direction in a kind of low-resolution image | |
CN109344898A (en) | Convolutional neural networks image classification method based on sparse coding pre-training | |
CN108268890A (en) | A kind of hyperspectral image classification method | |
CN105631478A (en) | Plant classification method based on sparse expression dictionary learning | |
Singh et al. | Leaf identification using feature extraction and neural network | |
Gao et al. | Natural scene recognition based on convolutional neural networks and deep Boltzmannn machines | |
CN114913379A (en) | Remote sensing image small sample scene classification method based on multi-task dynamic contrast learning | |
CN115222998A (en) | Image classification method | |
CN112800927A (en) | AM-Softmax loss-based butterfly image fine granularity identification method | |
CN112766283A (en) | Two-phase flow pattern identification method based on multi-scale convolution network | |
CN104376312B (en) | Face identification method based on bag of words compressed sensing feature extraction | |
Manzari et al. | A robust network for embedded traffic sign recognition |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180828 |
|
WD01 | Invention patent application deemed withdrawn after publication |