CN108564166A - Based on the semi-supervised feature learning method of the convolutional neural networks with symmetrical parallel link - Google Patents
Based on the semi-supervised feature learning method of the convolutional neural networks with symmetrical parallel link Download PDFInfo
- Publication number
- CN108564166A CN108564166A CN201810238288.4A CN201810238288A CN108564166A CN 108564166 A CN108564166 A CN 108564166A CN 201810238288 A CN201810238288 A CN 201810238288A CN 108564166 A CN108564166 A CN 108564166A
- Authority
- CN
- China
- Prior art keywords
- image
- network
- layer
- category
- training
- 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
- 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
Abstract
The invention discloses based on the semi-supervised feature learning method of the convolutional neural networks with symmetrical parallel link, comprise the following steps:It generates without category damaged image data set;Build parallel link convolutional neural networks;Pre-training image restores neural network;It extracts network parameter and builds sorter network;Training sorter network.The present invention utilizes the recovery tasks to no category image data, pre-training neural network to realize semi-supervised feature learning to improve to there is the classifying quality of class logo image.In addition, by the way that symmetrical parallel link is added in traditional convolutional neural networks autocoder, so that network is more easy to optimize, and enhance network middle level features abstracting power so that the network weight that unsupervised image recovery tasks obtain is more easy to migrate in supervised learning task.The present invention realizes efficient, the accurate semi-supervised learning method based on convolutional neural networks, therefore has higher practical value.
Description
Technical field
The present invention relates to the semi-supervised feature learnings of image, more particularly to based on the convolutional neural networks with symmetrical parallel link
(Convolutional Neural Network, CNN) semi-supervised feature learning method.
Background technology
With the continuous rapid development of information technology, every field daily all generate with surprising rapidity it is various types of
Image data.In a large amount of image data acquisition, communication process, image, semantic information how is more fully understood, and complete whereby
It is the significant challenge of artificial intelligence and area of pattern recognition now at the achievable task of the mankind's.People urgently wish to count
Calculation machine can help the mankind preferably to obtain and utilize mass image data.
Image data in internet be all often in the form of not category existing for, only a small amount of structural data or
Image data for scientific research has category.Therefore a large amount of unlabeled datas, auxiliary how to be used to have the reason of category data on a small quantity
Solution and study, become artificial intelligence field urgent problem to be solved.The semi-supervised feature learning of image is used as and utilizes no category number
According to important method, be constantly subjected to the extensive concern of industrial quarters and academia, and through frequently as various image correlation International Academics
The important theme of meeting is one very important research topic of artificial intelligence and area of pattern recognition.Its basic thought is profit
Without category information and there will be category data characteristics using certain technological means with the structural information extracted in no class logo image
It is associated, to which auxiliary has the understanding and study of class logo image.
In recent years, it is based on deep neural network, the method for especially depth convolutional neural networks is widely used in many meters
Calculation machine vision and pattern recognition task, many high-rise image understanding tasks, such as image classification, image segmentation the problems such as on obtain
The effect to attract people's attention.But it still has some disadvantages that it is made to receive limitation in the application, it is important that wherein it is needs
Largely there is category image data.There is category data volume limited, the performance of deep learning method is often not to the utmost such as people
Meaning.How by semi-supervised feature learning thought be applied to deep learning field, it has also become current research hot spot, to actively pushing forward society
Meeting IT application process plays an important role.While creating the social value that can not be substituted, still there are many crucial in the field
Technical problem not yet solves, and still there are many functions to realize that needs are further perfect, therefore, how to utilize depth convolutional Neural net
Network more effectively understands image in the case that semi-supervised, to realize the research of computer vision for greater flexibility, has deep
Remote meaning.
Invention content
Goal of the invention:It is a kind of based on band the technical problem to be solved by the present invention is in view of the deficiencies of the prior art, provide
The semi-supervised feature learning method of the convolutional neural networks of parallel link, by magnanimity unlabeled data to convolutional Neural net
Network carries out pre-training, final to improve the performance in having category data.
In order to solve the above-mentioned technical problem, the invention discloses a kind of based on the convolutional neural networks with parallel link
The semi-supervised feature learning method of (Convolutional Neural Network, CNN), comprises the following steps:
Step 1, it generates without category and has category data set:Acquisition has category and without category image data, to every image
Random cropping and normalized are done, category image collection X has been obtained0It is big according to image resolution ratio with no category image collection Y
It is small, different destructions is carried out to image in set Y, after being destroyed without category image set X1If Z is to have class logo image category
Vector, Z={ z1,z2,…,zn, ziIndicate that i-th image category, i values are 1~n, then (X1, Y) and it forms for non-supervisory pre-
It is trained without category training data set, (X0, Z) and as Training there is category training data set;
Step 2, structure pre-training image restores network:Image instauration net network is built according to input picture size, if network
Total depth is D layers, and D is even number, wherein first D/2 layers are convolutional layer, latter D/2 layers is warp lamination, and convolution kernel size is taken as 3x3,
Step-length is 1 or 2, and step change rate is determined according to network depth and image size.Input is the image set X after being destroyed in step 11
In image, export as image after network recovery;
Step 3, training image restores network:Use ADAM (Kingma, Diederik P., and Jimmy Ba. "
Adam:A method for stochastic optimization."arXiv preprint arXiv:1412.6980
(2014)) optimization algorithm, the training set (X obtained using step 11, Y) network constructed by step 2 is trained, with set X1
In be damaged image as input, and corresponding lossless image is used as network monitoring information using in set Y, and image is recorded after trained
D/2 layers of each layer of weight W and biasing b before recovery network;
Step 4, Supervised classification network is built, the image built using step 2 restores network as template, according to input picture
Size builds D/2 layer networks, is convolutional layer, and step change is consistent with network is built in step 2.And Max-pooling is added
Layer and Softmax layer, while to convolution layer parameter using trained in step 3 network respective weights W and biasing b carry out initially
Change;
Step 5, the sorter network for building and initializing in step 4 is used ADAM optimization algorithms by training sorter network,
Have and be trained in category image data, until algorithmic statement.
Step 1 specifically comprises the following steps:
Step 1-1, acquisition have category and without category image data, are cut to each image, using random cropping
Mode obtains the identical image block of size, and wherein tile size depends on original image size and model size, for being less than
The low-resolution image (such as CIFAR-10 data sets) of 50*50, cutting size is 29*29, for being more than 225*225 high-resolution
Natural image (such as PASCAL VOC data sets), cutting size be 225*225 first carried out if resolution ratio is between
Close resolution ratio is zoomed to, then is cut.All image collections after cutting are denoted as X ';
Image block after cutting is normalized and centralization processing, first image data after calculating cutting step 1-2
The mean value and standard deviation of each pixel on set X ', if all image mean value images on X ' areStandard deviation is std, for one
Specific image x is opened, it is normalized and centralization processing is as follows:
X ' is image x treated images;After processing in image, there is category image collection to be denoted as X0, no category image set
Conjunction is denoted as Y.
Step 1-3 is processed to rear image collection X for there is class logo image0There is category with corresponding category vector Z composition
Training data (X0, Z), Z={ z1,z2,…,zn, ziIndicate i-th image category.
Step 1-4 destroys the image in no category image collection Y, adds Gaussian noise or by pixel in image
Value is set to 0, if being low-resolution image (resolution ratio is less than 50*50) after cutting, takes and increases this Noise Method, if after cutting
For high-definition picture (resolution ratio is more than or equal to 50*50), then pixel is taken to be set to 0 method, the pixel set to 0 is random choosing
Take 10 adjacent regions 8*8, after being destroyed without category image set X1, formed without category with no category image collection Y
Training data set (X1,Y)。
Step 2 specifically comprises the following steps:
Step 2-1, if it is D layers that image, which restores network total depth, D is even number, wherein first D/2 layers use convolutional layer, rear D/2
It is 3x3 that layer, which uses warp lamination, convolution kernel size, and step-length is 1 or 2, every k layers of step-length be+1 layer of step-length of 1 kth is 2,0<k<D/
2-1 repeats n times.K and n sizes are determined by network depth and tile size:For 29*29 low-resolution images, k=4, n=
3, for 225*225 high-definition pictures, k=2, n=5 are added after every layer of convolutional layer and warp lamination
BatchNormalization layers and ReLU (Rectified Linear Unit) non-linear layer, network inputs are raw in step 1
At destruction after image collection X1Middle image exports as image after network recovery.Network parameter includes convolutional layer and deconvolution
B and BatchNormalization layers of weight γ of weight W and biasing and biasing β of layer.
Symmetrical parallel link is added every two layers in step 2-2 between convolutional layer and warp lamination:If COiIndicate i-th
Layer convolutional layer output, DIiRepresent i-th layer of warp lamination input, DOiI-th layer of warp lamination output is represented, then parallel link indicates
For:
DID-i+1=DOD-i+COi,
Then DID+1It is exported for network, CO0For network inputs, first time parallel link is initially connected to export from input layer
Layer, the input of network end layer calculate the loss of Euclidean distance network training as after with correspondence image in lossless image set Y
Function
Wherein XiFor image set X1In i-th image, YiFor i-th image in image set Y,It is represented for neural network
Function, N be training image quantity, θ be network it is all can training parameter, include the weight W and biasing of convolutional layer and warp lamination
B and BatchNormalization layers of weight γ and biasing β.
Step 3 specifically comprises the following steps:
Step 3-1 carries out gradient backpropagation using ADAM optimization algorithms and neural network, learning rate is trained to be set as 1e-
4, training continues n1Wheel (generally 20 wheels), n-th2Wheel (generally the 8th wheel) and n-th3After wheel (generally the 16th wheel),
Learning rate is respectively set to 1e-5 and 1e-6;;
Step 3-2, to illustrate the gradient backpropagation step with parallel link, if it is 7 layers that image, which restores network depth, with
Mode adds parallel link in step 2-2, if X0For network inputs, XiIt is exported for i-th layer of convolutional layer, cross-layer links specifically by X1
It is connected to the 5th layer of input, by X0It is connected to the 7th layer of input.At this time in forward calculation, obtains image and restore network output X7For:
X7=f7(X0, X6);
Step 3-3, by X7It further spreads out and is expressed as:
X7=f7(X0, X6)
=f7(X0, f6(X5))
=f7(X0, f6(f5(X1, X4)))
=f7(X0, f6(f5(X1, f4(X3))))
=f7(X0, f6(f5(X1, f4(fk(X2)))))
Step 3-4, in gradient backpropagation, i-th layer in network directly obtains gradient to update the layer from its top layer
Parameter θi, in the method, θiSpecifically include convolution/warp lamination weight WiWith biasing bi, and
BatchNormalization layers of weight γiWith biasing βiBy taking the network first tier as an example, to update first layer parameter θ1It needs to count
Loss function ζ is calculated about θ1Partial derivative:
Step 3-5, after the partial derivative for obtaining each layer of corresponding top layer, more using the corresponding update rule of ADAM algorithms
New each layer parameter, training step 1 obtain without category training data set (X1, Y) on carry out, with X1In be damaged image
As input, clear image is corresponded to using in Y as supervision message with method undated parameter in step 3-4, is owned until having used
Stop after 20 wheel of training data training.
Step 4 specifically comprises the following steps:
Build Supervised classification network, train in first extraction step 3 image restore network convolutional layer parameter W, b, γ and
β builds D/2 layer networks according to input picture size, and every layer is convolutional layer, and step change is consistent with network is built in step 2,
Maximum value pond layer (Max-pooling) layer is added wherein after last layer of convolutional layer, behind according to classification task category
Quantity N, N have category training data set (X for Training0, Z) in category vector Z maximum possible value, be added N
The Softmax layers of class.Using parameter W, b, γ and β of extraction, by the way of indirect assignment, Supervised classification network is corresponded to
Parameter initialization.
Step 5 specifically comprises the following steps:
Training sorter network, is having category by the sorter network for building and initializing in step 4 using ADAM optimization algorithms
Training data set (X0, Z) on be trained, initial learning rate is set as 1e-4, and training continues n4Wheel (generally 200 wheels),
N-th5、n6And n7Take turns (general n5For 80, n6For 120, n7After 160), current learning rate is multiplied by 0.2 and is newly learnt
Rate, n4Until network convergence after wheel.
The present invention is directed to the depth convolutional neural networks method of the semi-supervised feature learning of image, and the present invention has following special
Sign:1) present invention is carrying out pre-training constantly using deep neural network, adds parallel link so that network can be received faster
It holds back, while network middle level features being allow to retain more image abstraction information;2) the method for the present invention is different from the past for spy
The semi-supervised feature learning method for determining data type, can be applied to nearly all image data, have universality.
Advantageous effect:The present invention has fully considered that convolutional neural networks are successively connected to non-supervisory feature learning, be added across
Layer connection ensures that network can extract abstracted information useful enough from unlabeled data, to which preferably auxiliary has category
The classification of data promotes the accuracy rate of image classification.
Description of the drawings
The present invention is done with reference to the accompanying drawings and detailed description and is further illustrated, it is of the invention above-mentioned or
Otherwise advantage will become apparent.
Fig. 1 is flow chart of the present invention.
Fig. 2 is schematic network structure.
Fig. 3 a are the secondary artwork in embodiment.
Fig. 3 b are image after Fig. 3 a plus noises.
Fig. 3 c are the recovery image of Fig. 3 a.
Fig. 4 a are the secondary artwork in embodiment.
Fig. 4 b are the secondary artwork in embodiment.
Fig. 4 c are the corresponding characteristic patterns of Fig. 4 a.
Fig. 4 d are the corresponding characteristic patterns of Fig. 4 b.
Specific implementation mode
The present invention will be further described with reference to the accompanying drawings and embodiments.
As shown in Figure 1, the invention discloses a kind of image repair method based on the convolutional neural networks with parallel link,
It comprises the following steps:
Step 1, it generates without category and has category data set:Acquisition has category and without category image data, to every image
Random cropping and normalized are done, category image set X has been obtained0With no category clear image set Y, according to image resolution ratio
Size carries out image in Y different destructions, the image set X without category after being destroyed1If Z be have class logo image category to
Amount, Z={ z1,z2,…,zn, ziIndicate i-th image category, then (X1, Y) and neural network of the composition for non-supervisory pre-training
Training set, (X0, Z) and training set as Training;
Step 2, structure pre-training image restores network:It is built according to input picture size, if network total depth is D layers, D
For even number, wherein first D/2 layers are convolutional layer, latter D/2 layers is warp lamination, and convolution kernel size is taken as 3x3, and step-length is 1 or 2, root
Step change rate is determined according to network depth and image size.Input is that image set X is destroyed in step 11Middle image exports as network
Image after recovery;
Step 3, training image restores network:Using ADAM, (adaptive moment estimation, adaptive square are estimated
Meter) optimization algorithm, the training set (X obtained using step 11, Y) network constructed by step 2 is trained, with set X1In by
Image is destroyed as input, and corresponding lossless image is as network monitoring information using in set Y, record image restores after training
D/2 layers of each layer of weight W and biasing b before network;
Step 4, Supervised classification network is built, the image built using step 2 restores network as template, according to input picture
Size builds D/2 layer networks, is convolutional layer, and step change is consistent with network is built in step 2.And Max-pooling is added
Layer and Softmax layer, while to convolution layer parameter using trained in step 3 network respective weights W and biasing b carry out initially
Change;
Step 5, the sorter network for building and initializing in step 4 is used ADAM optimization algorithms by training sorter network,
Have and be trained in category image data, until algorithmic statement.
Step 1 specifically comprises the following steps:
This step describes process of data preprocessing, and acquisition has category and without category image data, cut out to each image
Cut, by the way of random cropping, obtain the identical image block of size, wherein tile size depend on original image size and
Model size, for the low-resolution image (such as CIFAR-10 data sets) less than 50*50, cutting size is 29*29, for big
In the high-resolution natural images of 225*225 (such as PASCAL VOC data sets), cutting size is 225*225, if resolution ratio exists
Therebetween, then it first carries out zooming to close resolution ratio, then is cut.All image collections after cutting are denoted as X ';It will cut out
Image block after cutting be normalized with centralization processing, calculate after cutting the equal of each pixel on sets of image data X ' first
Value and standard deviation, if all image mean value images on X ' areStandard deviation is std, for a specific image x, is carried out to it
Normalization and centralization processing are as follows:
X ' is image x treated images;After processing in image, there is category image collection to be denoted as X0, no category image set
Conjunction is denoted as Y.
For there is class logo image, it is processed to rear image collection X0There is category training data with corresponding category vector Z composition
(X0, Z), Z={ z1,z2,…,zn, ziIndicate i-th image category.
It for the image in no category clear image set Y, is destroyed, add Gaussian noise or is set pixel value in image
It is 0, if being low-resolution image (resolution ratio is less than 50*50) after cutting, takes and increase this Noise Method, if being high after cutting
Image in different resolution (resolution ratio is more than or equal to 50*50), then take pixel to be set to 0 method, and the pixel set to 0 is to randomly select 10
A adjacent regions 8*8.After being destroyed without category image set X1, with clear image collection Y compositions without category training data
(X1,Y)。
Step 2 specifically comprises the following steps:
This step describes the building process to pre-training neural network model, if network total depth is D layers, D is even number,
In first D/2 layers use convolutional layer, latter D/2 layer use warp lamination, convolution kernel size be 3x3, step-length be 1 or 2, every k layers walk
A length of+1 layer of step-length of 1 kth is 2,0<k<D/2-1 repeats n times.K and n sizes are adjusted according to network depth and tile size.
BatchNormalization layers and ReLU (Rectified Linear Unit) are added after every layer of convolutional layer and warp lamination
Non-linear layer.Network inputs are the image after the destruction generated in step 1, are exported as image after network recovery;Every two layers,
Symmetrical parallel link is added between convolutional layer and warp lamination.Specifically, if COiIndicate i-th layer of convolutional layer output, DIiRepresent
I layers of warp lamination input, DOiI-th layer of warp lamination output is represented, then parallel link is represented by:
DID-i+1=DOD-i+COi
Particularly, DID+1It is exported for network, CO0For network inputs, i.e. first time parallel link is connected since input layer
To output layer.Network end layer is inputted calculates Euclidean distance as loss function with correspondence image in original digital image data collection Y:
Wherein XiFor no category damaged image collection X1In i-th image, YiFor i-th figure in no category clear image collection Y
Picture,For neural network represent function, N be training image quantity, θ be network it is all can training parameter, including convolutional layer and
B and BatchNormalizatiion layers of weight γ of weight W and biasing and biasing β of warp lamination.
Fig. 2 is network structure rough schematic, and in the figure of the left side, Corrupted data are network inputs data,
Restored data are network output datas, and conv1, conv2 and c3 ... c6 are convolutional layer, d3 ... d6, deconv1,
Deconv2 is warp lamination.The right figure describes the details of a parallel link, and conv is convolutional layer in figure, and deconv is warp
Lamination, ReLU and BatchNorm respectively represent ReLU layers and BatchNormalization layers.
Step 3 specifically comprises the following steps:
This step describes the training process to pre-training neural network model, and it is reversed to carry out gradient using ADAM optimization algorithms
Training neural network is propagated, learning rate is set as 1e-4.Training continues 20 wheels, after the 8th wheel and the 16th wheel, learning rate point
1e-5 and 1e-6 are not set as it.
To illustrate the gradient backpropagation step with parallel link, if the network depth is 7 layers, in a manner of in step 2-2
Parallel link is added, if X0For network inputs, XiIt is exported for i-th layer of convolutional layer, parallel link is specifically by X1Be connected to the 5th layer it is defeated
Enter, by X0It is connected to the 7th layer of input.At this time in forward calculation, obtains image and restore network output X7For:
X7=f7(X0, X6)
X7It can be further represented as:
X7=f7(X0, X6)
=f7(X0, f6(X5))
=f7(X0, f6(f5(X1, X4)))
=f7(X0, f6(f5(X1, f4(X3))))
=f7(X0, f6(f5(X1, f4(fk(X2)))))
Wherein X1And X2Indicate the characteristic pattern that the 1st, 2 convolutional layer obtains.
In gradient backpropagation, i-th layer in network directly obtains gradient to update the parameter θ of this layer from its top layeri,
In the method, θiSpecifically include convolution/warp lamination weight WiWith biasing biAnd BatchNormalization layers of weight
γiWith biasing βiBy taking the network first tier as an example, to update first layer parameter θ1Need counting loss function ζ about θ1Local derviation
Number:
After the partial derivative for obtaining each layer of corresponding top layer, the corresponding each layer of update Policy Updates of ADAM algorithms is used
Parameter, the unlabeled data set (X that training is obtained in step 11, Y) on carry out, with X1In be damaged image as input, with Y
Middle corresponding clear image as supervision message with method undated parameter in step 3-4, until having used all training datas to train
Stop after 20 wheels.
Step 4 specifically comprises the following steps:
This step describes the building process of Supervised classification network, and image is trained in first extraction step 3 and restores network volume
Lamination parameter W, b, γ and β build D/2 layer networks according to input picture size, and every layer is convolutional layer, step change and step
It is consistent that network is built in 2, wherein maximum value pond layer (Max-pooling) layer is added after last layer of convolutional layer, behind
It is Training data (X according to classification task category quantity N, N0, Z) in category vector Z maximum possible value, be added N
The Softmax layers of class.Using parameter W, b, γ and β of extraction, by the way of indirect assignment, Supervised classification network is corresponded to
Parameter initialization.
Step 5 specifically comprises the following steps:
This step describes the training process of Supervised classification network, will be built in step 4 and the sorter network initialized makes
There is category image data set (X with ADAM optimization algorithms0, Z) on be trained, initial learning rate is set as 1e-4, and training is held
Continuous n4Wheel (generally 200 wheels), n-th5、n6And n7Take turns (general n5For 80, n6For 120, n7After 160), will currently it learn
Habit rate is multiplied by 0.2 and obtains new learning rate, n4Until network convergence after wheel.
Embodiment 1
The present embodiment describes the semi-supervised feature learning on CIFAR-10, including with lower part:
1,50000 natural images in CIFAR-10 data sets are evenly divided into two parts first, a part includes
It includes 46000 without class logo image that 4000, which have class logo image, another part,.
2, for every 32*32 size without class logo image, in training, the random image block for intercepting 29*29, in image
Upper additional mean value is 0, the Gaussian noise that standard deviation is 30.For after plus noise image and the image of plus noise does not carry out respectively
Normalization is formed without category training set.
3,18 layers of convolutional neural networks with parallel link are built, it is enterprising without class logo image in generation using ADAM algorithms
Row training.Retain first 9 layers after network convergence, corresponding sorter network is built using its network weight.
4, there is training sorter network in class logo image at other 4000, using the training of ADAM algorithms until convergence, in original
It is tested on 50000 test sets of beginning image, and reports accuracy rate such as the following table 1:
Table 1
Wherein last column is this method accuracy rate, it can be seen that this method reaches half carried out using GAN many at present
Supervised learning accuracy rate and compared to without using pre-training (No pre-training rows) and being not added with parallel link (Pre-
Training without shortcut rows) identical network, accuracy rate has a distinct increment.
Embodiment 2
The present embodiment description utilizes the extensive semi-supervised feature of 2007 data of Imagenet data sets and Pascal VOC
Study, including with lower part:
1, first on Imagenet natural image data sets, the random image block for intercepting 225*225, for each image
Its random 35 8*8 image-region pixel is set to 0 by block, and the image and original image that set to 0 processing are normalized respectively,
It is formed without category training set.
2,32 layers of convolutional neural networks with parallel link are built, it is enterprising without class logo image in generation using ADAM algorithms
Row training.Retain first 16 layers after network convergence, corresponding sorter network is built using its network weight.
3, to 2007 natural image data of PASCAL VOC, the random image block for intercepting 225*225, and by image block with
After 50% probability carries out flip horizontal and normalizes, category data have been obtained.
4, there is training sorter network in category data in generation, using the training of ADAM algorithms until convergence, use test
Collection is tested, and reports accurate final really rate such as following table:
Table 2
Wherein last column is this method accuracy rate, it can be seen that this method higher than current congenic method accuracy rate 1% is left
The right side, and compared to parallel link (Ours without shortcut rows) identical network is not added with, accuracy rate has a distinct increment.Figure
3a~Fig. 3 c are recovery effects of the embodiment pre-training network to image, and Fig. 3 a are artwork, and Fig. 3 b are image after plus noise, figure
3c is to restore image, it can be seen that pre-training network can learn well to image detail information.Fig. 4 a~Fig. 4 d are the realities
The feature visualization effect that example learns is applied, Fig. 4 a and Fig. 4 b are artwork, and Fig. 4 c and Fig. 4 d are respectively Fig. 4 a and Fig. 4 b corresponding
Characteristic pattern, the dog face part in each image are all fairly obvious in characteristic pattern, it may be said that the feature that bright party's calligraphy learning arrives
Image Deep Semantics information is captured well.
The present invention provides based on the semi-supervised feature learning method of the convolutional neural networks with symmetrical parallel link, specific reality
Now there are many method of the technical solution and approach, the above is only a preferred embodiment of the present invention, it is noted that for this
For the those of ordinary skill of technical field, without departing from the principle of the present invention, several improvement and profit can also be made
Decorations, these improvements and modifications also should be regarded as protection scope of the present invention.Each component part being not known in the present embodiment is available
The prior art is realized.
Claims (6)
1. based on the semi-supervised feature learning method of the convolutional neural networks with symmetrical parallel link, which is characterized in that including as follows
Step:
Step 1, it generates without category and has category data set:Acquisition has category and without category image data, to every image do with
Machine is cut and normalized, has obtained category image collection X0With no category image collection Y, according to image resolution ratio size,
Different destructions is carried out to image in set Y, after being destroyed without category image set X1If Z be have class logo image category to
Amount, Z={ z1,z2,…,zn, ziIndicate that i-th image category, i values are 1~n, then (X1, Y) and it forms for non-supervisory pre- instruction
It is experienced without category training data set, (X0, Z) and as Training there is category training data set;
Step 2, structure pre-training image restores network:Image instauration net network is built according to input picture size, if network aggregate depth
Degree is D layers, and D is even number, wherein first D/2 layers are convolutional layer, latter D/2 layers is warp lamination, and convolution kernel size is taken as 3x3, is inputted
For the image set X after destruction in step 11In image, export as image after network recovery;
Step 3, training image restores network:Using ADAM optimization algorithms, the training set (X obtained using step 11, Y) and to step 2
Constructed network is trained, with set X1In be damaged image as input, and using in set Y corresponding lossless image as
Network monitoring information, record image restores D/2 layers of each layer of weight W and biasing b before network after training;
Step 4, Supervised classification network is built:The image built using step 2 restores network as template, according to input picture size
D/2 layer networks are built, are convolutional layer, and Max-pooling layers and Softmax layers are added, while convolution layer parameter is used
Network respective weights W is trained in step 3 and biasing b is initialized;
Step 5, the sorter network for building and initializing in step 4 is used ADAM optimization algorithms, is there is class by training sorter network
It is trained in logo image data, until algorithmic statement.
2. according to the method described in claim 1, it is characterized in that, step 1 includes the following steps:
Step 1-1, acquisition has category and without category image data, is cut to each image, by the way of random cropping,
The identical image block of size is obtained, wherein tile size depends on original image size and model size, will own after cutting
Image collection is denoted as X ';
Image block after cutting is normalized and centralization processing, first sets of image data after calculating cutting step 1-2
The mean value and standard deviation of each pixel on X ', if all image mean value images on X ' areStandard deviation is std, for Yi Zhangte
Determine image x, it is normalized and centralization processing is as follows:
X ' is image x treated images, in image after treatment, will have category image collection to be denoted as X0, no category image set
Conjunction is denoted as Y;
Step 1-3 is processed to rear image collection X for there is class logo image0There is category to train number with corresponding category vector Z composition
According to (X0, Z), Z={ z1,z2,…,zn, ziIndicate i-th image category;
Step 1-4 destroys the image in no category image collection Y, adds Gaussian noise or sets pixel value in image
It is 0, if being low-resolution image after cutting, takes and increase this Noise Method, if being high-definition picture after cutting, takes
Pixel is set to 0 method, and the pixel set to 0 is to randomly select 10 adjacent regions 8*8, after being destroyed without class logo image
Collect X1, formed without category training data set (X with no category image collection Y1,Y)。
3. according to the method described in claim 2, it is characterized in that, step 2 includes the following steps:
Step 2-1, if it is D layers that image, which restores network total depth, D is even number, wherein first D/2 layers use convolutional layer, latter D/2 layers makes
With warp lamination, convolution kernel size is 3x3, and step-length is 1 or 2, and after k layers of step-length are 1 ,+1 layer of step-length of kth is 2, repeats n times, k tools
Body value is 0<k<D/2 adjusts k and n sizes according to hands-on image size, if after cutting being low-resolution image, k=4,
N=3;If after cutting being high-definition picture, k=2, n=5;It is added after every layer of convolutional layer and warp lamination
BatchNormalization layers and ReLU non-linear layers, network inputs are the image collection X after the destruction generated in step 11
Middle image exports as image after network recovery;Network parameter includes the weight W and biasing b of convolutional layer and warp lamination, and
BatchNormalization layers of weight γ and biasing β;
Symmetrical parallel link is added every two layers in step 2-2 between convolutional layer and warp lamination:If COiIndicate i-th layer of convolution
Layer output, DIiRepresent i-th layer of warp lamination input, DOiI-th layer of warp lamination output is represented, then parallel link is expressed as:
DID-i+1=DOD-i+COi,
Then DID+1It is exported for network, CO0For network inputs, first time parallel link is initially connected to output layer, network from input layer
End layer inputs the loss function that Euclidean distance network training as after is calculated with correspondence image in lossless image set Y
Wherein XiFor image set X1In i-th image, YiFor i-th image in image set Y,For neural network represent function,
N be training image quantity, θ be network it is all can training parameter, including the weight W of convolutional layer and warp lamination and biasing b, and
BatchNormalization layers of weight γ and biasing β.
4. according to the method described in claim 3, it is characterized in that, step 3 includes the following steps:
Step 3-1 carries out gradient backpropagation using ADAM optimization algorithms and neural network, learning rate is trained to be set as 1e-4, instructs
Practice and continues n1Wheel, n-th2Wheel and n-th3After wheel, learning rate is respectively set to 1e-5 and 1e-6;
Step 3-2 adds parallel link, if X if it is 7 layers that image, which restores network depth, in a manner of in step 2-20It is defeated for network
Enter, XiIt is exported for i-th layer of convolutional layer, cross-layer links specifically by X1It is connected to the 5th layer of input, by X0It is connected to the 7th layer of input, this
When in forward calculation, obtain image restore network output X7For:
X7=f7(X0, X6);
Step 3-3, by X7It further spreads out and is expressed as:
X7=f7(X0, X6)
=f7(X0, f6(X5))
=f7(X0, f6(f5(X1, X4)))
=f7(X0, f6(f5(X1, f4(X3))))
=f7(X0, f6(f5(X1, f4(fk(X2)))));
Step 3-4, in gradient backpropagation, i-th layer in network directly obtains gradient to update the ginseng of this layer from its top layer
Number θi;
Step 3-5, it is every using the corresponding update Policy Updates of ADAM algorithms after the partial derivative for obtaining each layer of corresponding top layer
One layer parameter, training step 1 obtain without category training data set (X1, Y) on carry out, with X1In be damaged image conduct
Input, using in Y correspondence image as supervision message with method undated parameter in step 3-4, until having used all training datas
Stop after 20 wheel of training.
5. method according to claim 4, which is characterized in that step 4 includes:
Supervised classification network is built, image is trained in first extraction step 3 and restores network convolutional layer parameter, according to input picture
Size builds D/2 layer networks, and every layer is convolutional layer, and step change is consistent with network is built in step 2, wherein in last layer
Layer Max-pooling layers of maximum value pond is added after convolutional layer, is behind to have supervision to instruct according to classification task category quantity N, N
Experienced has category training data set (X0, Z) in category vector Z maximum possible value, be added N classes Softmax layers, use
Supervised classification network is corresponded to parameter initialization by the parameter of extraction by the way of indirect assignment.
6. method according to claim 5, which is characterized in that step 5 includes the following steps:
Training sorter network, is having category training by the sorter network for building and initializing in step 4 using ADAM optimization algorithms
Data acquisition system (X0, Z) on be trained, initial learning rate is set as 1e-4, and training continues n4Wheel, n-th5、n6And n7Wheel terminates
Afterwards, current learning rate is multiplied by 0.2 and obtains new learning rate, n4Until network convergence after wheel.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810238288.4A CN108564166A (en) | 2018-03-22 | 2018-03-22 | Based on the semi-supervised feature learning method of the convolutional neural networks with symmetrical parallel link |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810238288.4A CN108564166A (en) | 2018-03-22 | 2018-03-22 | Based on the semi-supervised feature learning method of the convolutional neural networks with symmetrical parallel link |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108564166A true CN108564166A (en) | 2018-09-21 |
Family
ID=63532070
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810238288.4A Pending CN108564166A (en) | 2018-03-22 | 2018-03-22 | Based on the semi-supervised feature learning method of the convolutional neural networks with symmetrical parallel link |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108564166A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110543563A (en) * | 2019-08-20 | 2019-12-06 | 暨南大学 | Hierarchical text classification method and system |
CN110598629A (en) * | 2019-09-11 | 2019-12-20 | 北京百度网讯科技有限公司 | Search space construction method and device of hyper network and electronic equipment |
WO2020061972A1 (en) * | 2018-09-27 | 2020-04-02 | 电子科技大学 | Convolutional neural network-based diabetic retinopathy diagnosing technique |
CN111263226A (en) * | 2020-01-17 | 2020-06-09 | 中国科学技术大学 | Video processing method, video processing device, electronic equipment and medium |
CN115147426A (en) * | 2022-09-06 | 2022-10-04 | 北京大学 | Model training and image segmentation method and system based on semi-supervised learning |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106204449A (en) * | 2016-07-06 | 2016-12-07 | 安徽工业大学 | A kind of single image super resolution ratio reconstruction method based on symmetrical degree of depth network |
AU2017101166A4 (en) * | 2017-08-25 | 2017-11-02 | Lai, Haodong MR | A Method For Real-Time Image Style Transfer Based On Conditional Generative Adversarial Networks |
CN107316307A (en) * | 2017-06-27 | 2017-11-03 | 北京工业大学 | A kind of Chinese medicine tongue image automatic segmentation method based on depth convolutional neural networks |
CN107784308A (en) * | 2017-10-09 | 2018-03-09 | 哈尔滨工业大学 | Conspicuousness object detection method based on the multiple dimensioned full convolutional network of chain type |
-
2018
- 2018-03-22 CN CN201810238288.4A patent/CN108564166A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106204449A (en) * | 2016-07-06 | 2016-12-07 | 安徽工业大学 | A kind of single image super resolution ratio reconstruction method based on symmetrical degree of depth network |
CN107316307A (en) * | 2017-06-27 | 2017-11-03 | 北京工业大学 | A kind of Chinese medicine tongue image automatic segmentation method based on depth convolutional neural networks |
AU2017101166A4 (en) * | 2017-08-25 | 2017-11-02 | Lai, Haodong MR | A Method For Real-Time Image Style Transfer Based On Conditional Generative Adversarial Networks |
CN107784308A (en) * | 2017-10-09 | 2018-03-09 | 哈尔滨工业大学 | Conspicuousness object detection method based on the multiple dimensioned full convolutional network of chain type |
Non-Patent Citations (8)
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020061972A1 (en) * | 2018-09-27 | 2020-04-02 | 电子科技大学 | Convolutional neural network-based diabetic retinopathy diagnosing technique |
CN110543563A (en) * | 2019-08-20 | 2019-12-06 | 暨南大学 | Hierarchical text classification method and system |
CN110543563B (en) * | 2019-08-20 | 2022-03-08 | 暨南大学 | Hierarchical text classification method and system |
CN110598629A (en) * | 2019-09-11 | 2019-12-20 | 北京百度网讯科技有限公司 | Search space construction method and device of hyper network and electronic equipment |
CN110598629B (en) * | 2019-09-11 | 2023-06-06 | 北京百度网讯科技有限公司 | Super-network search space construction method and device and electronic equipment |
CN111263226A (en) * | 2020-01-17 | 2020-06-09 | 中国科学技术大学 | Video processing method, video processing device, electronic equipment and medium |
CN111263226B (en) * | 2020-01-17 | 2021-10-22 | 中国科学技术大学 | Video processing method, video processing device, electronic equipment and medium |
CN115147426A (en) * | 2022-09-06 | 2022-10-04 | 北京大学 | Model training and image segmentation method and system based on semi-supervised learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108564166A (en) | Based on the semi-supervised feature learning method of the convolutional neural networks with symmetrical parallel link | |
Jia et al. | Focnet: A fractional optimal control network for image denoising | |
CN105512289B (en) | Image search method based on deep learning and Hash | |
CN104113789B (en) | On-line video abstraction generation method based on depth learning | |
Fu et al. | Fast crowd density estimation with convolutional neural networks | |
CN107358257B (en) | Under a kind of big data scene can incremental learning image classification training method | |
Wang et al. | I3d-lstm: A new model for human action recognition | |
CN107862668A (en) | A kind of cultural relic images restored method based on GNN | |
CN106355248A (en) | Deep convolution neural network training method and device | |
CN104778448B (en) | A kind of face identification method based on structure adaptive convolutional neural networks | |
CN103824309B (en) | Automatic extracting method of urban built-up area border | |
CN106228512A (en) | Based on learning rate adaptive convolutional neural networks image super-resolution rebuilding method | |
CN106295584A (en) | Depth migration study is in the recognition methods of crowd's attribute | |
CN108280233A (en) | A kind of VideoGIS data retrieval method based on deep learning | |
CN108734095A (en) | A kind of motion detection method based on 3D convolutional neural networks | |
CN107330845A (en) | A kind of digital picture steganalysis method based on convolutional neural networks | |
Xu et al. | Recurrent convolutional neural network for video classification | |
CN108537747A (en) | A kind of image repair method based on the convolutional neural networks with symmetrical parallel link | |
CN107247991A (en) | A kind of method and device for building neutral net | |
CN110751072B (en) | Double-person interactive identification method based on knowledge embedded graph convolution network | |
CN107423747A (en) | A kind of conspicuousness object detection method based on depth convolutional network | |
CN110490265A (en) | A kind of image latent writing analysis method based on two-way convolution sum Fusion Features | |
CN107784360A (en) | Step-by-step movement convolutional neural networks beta pruning compression method | |
CN108829810A (en) | File classification method towards healthy public sentiment | |
CN107480723A (en) | Texture Recognition based on partial binary threshold learning network |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180921 |
|
RJ01 | Rejection of invention patent application after publication |