CN118114723A - Convolutional neural network training method based on asymmetric convolution kernel - Google Patents
Convolutional neural network training method based on asymmetric convolution kernel Download PDFInfo
- Publication number
- CN118114723A CN118114723A CN202410383952.XA CN202410383952A CN118114723A CN 118114723 A CN118114723 A CN 118114723A CN 202410383952 A CN202410383952 A CN 202410383952A CN 118114723 A CN118114723 A CN 118114723A
- Authority
- CN
- China
- Prior art keywords
- neural network
- convolutional neural
- convolution
- data
- batch
- 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
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 59
- 238000012549 training Methods 0.000 title claims abstract description 32
- 238000000034 method Methods 0.000 title claims abstract description 25
- 238000010606 normalization Methods 0.000 claims abstract description 14
- 238000013528 artificial neural network Methods 0.000 claims abstract description 5
- 230000008569 process Effects 0.000 claims abstract description 5
- 238000012795 verification Methods 0.000 claims abstract description 4
- 238000012360 testing method Methods 0.000 claims description 9
- 238000011156 evaluation Methods 0.000 claims description 8
- 238000005457 optimization Methods 0.000 description 6
- 230000006399 behavior Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011423 initialization method Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a convolutional neural network training method based on an asymmetric convolutional kernel, which comprises the following steps: s01: acquiring a data set; s02: splitting n-n convolution kernels of a designated convolution layer in a convolution neural network into n-1 and 1*n asymmetric convolution kernels to form two parallel convolution layers, and adding a batch normalization layer after the split two parallel convolution layers; s03: training the convolutional neural network processed in the step S02 by adopting the data set acquired in the step S01, acquiring an image feature map in the training process, extracting image features, and classifying images; s04: after the convolutional neural network is trained in one round, the convolutional neural network is verified by using a verification set; s05: and (3) circularly executing the steps S2 to S4, and performing multi-round training on the convolutional neural network until reaching the ending condition of model convergence of the convolutional neural network. The method can reduce the number of network parameters on the premise of ensuring the network classification performance in the classification task.
Description
Technical Field
The invention relates to a convolutional neural network training method based on an asymmetric convolutional kernel.
Background
At present, a convolutional neural network (Convolutional Neural Networks, CNN) is a classical model in deep learning, the CNN mainly comprises a convolutional layer, a downsampling layer and a full-connection layer, and the CNN can gradually and deeply extract rich high-level semantics of an image by combining a plurality of convolutional layers and the downsampling layer, which is key information for completing tasks such as image classification or recognition.
The chinese patent with publication No. CN114091648a discloses a convolutional neural network-based image classification method and apparatus, and a convolutional neural network, in which the number of extracted feature values is increased by setting parallel sub-convolutional neural network layers, although the convolutional neural network is excellent in computer vision tasks, as a network model becomes larger, a network of up to several hundred million parameters may be used sometimes, which leads to a significant increase in memory occupation and computation requirements, so that restrictions on memory and computation resources need to be considered.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a convolutional neural network training method based on an asymmetric convolutional kernel, which can reduce the number of network parameters on the premise of ensuring the network classification performance in a classification task.
In order to solve the technical problems, the technical scheme of the invention is as follows: a convolutional neural network training method based on an asymmetric convolutional kernel comprises the following steps:
s01: acquiring a data set, wherein the data set is used for training a convolutional neural network;
s02: splitting n-n convolution kernels of a designated convolution layer in a convolution neural network into n-1 and 1*n asymmetric convolution kernels to form two parallel convolution layers, and adding a batch normalization layer after the split two parallel convolution layers;
s03: training the convolutional neural network processed in the step S02 by adopting the data set acquired in the step S01, acquiring an image feature map in the training process, extracting image features, and classifying images;
s04: after the convolutional neural network is trained in one round, the convolutional neural network is verified by using a verification set;
s05: and (3) circularly executing the steps S2 to S4, and performing multi-round training on the convolutional neural network until reaching the ending condition of model convergence of the convolutional neural network.
Further, in order to obtain a larger-scale data set to be applied to the neural network model test, in step S01, data enhancement is performed on the data set.
Further, the data enhancement mode is as follows: at least one of blurring and rotation.
Further, in order to ensure that the size of the output is consistent with that of the normal convolution, in step S2, after splitting the n×n convolution kernels of the specified convolution layers into two asymmetric convolution kernels, a padding operation is added.
Further, the filling formula of the filling operation is:
wherein,
W, H is the width and height of the pattern;
KW, KH is the convolution kernel width and convolution kernel height;
PW, PH is left and right filling size and up and down filling size;
s is convolution step length; w N,HN is the updated pattern width and pattern height.
Further, BN formula for batch normalization layer is:
yi＝BNγ,β(xi)；
wherein,
I is an index number in the batch data;
x i is the ith sample input data in the batch data;
y i is the i-th sample output data;
gamma, beta are self-learning parameters;
m is the total number of index batch input samples for a sample in a batch of input samples.
Mu Β is the mean of the batch input data;
variance of input data for a batch;
Is normalized data;
E is a very small number to avoid dividing by 0.
Further, the method further comprises the following steps:
S06: calculating the accuracy and positive and negative sample prediction bias values of the convolutional neural network on the test set, wherein the convolutional neural network uses the accuracy and the positive and negative sample prediction bias values together as evaluation indexes, and the final algorithm target of the convolutional neural network is as follows: the positive and negative samples have the lowest predicted bias value.
After the technical scheme is adopted, the invention has the following beneficial effects:
Firstly, according to the convolutional neural network training method based on the asymmetric convolutional kernels, the shape of the convolutional kernels is changed, one n-by-n convolutional kernel is decomposed into 1*n and n-by-1 convolutional kernels, the two kernels are used in parallel, the number of parameters of a model can be obviously reduced by introducing the technology, meanwhile, in order to avoid the problem caused by adding the asymmetric convolutional, a batch normalization layer is added after two split parallel convolutional layers, the training speed is improved, and finally, the classification performance of pictures is equivalent to that of a traditional convolutional neural network, and the number of network parameters is reduced.
Secondly, the method aims at the problem that the accuracy of the multi-classification evaluation indexes is similar, adopts the positive and negative sample deviation evaluation indexes, and uses the accuracy and the positive and negative sample prediction deviation value together as the evaluation indexes, so that the evaluation result is more reasonable.
Drawings
FIG. 1 is a diagram of a ResNet network architecture based on asymmetric convolution optimization in accordance with the present invention;
FIG. 2 is a diagram of a batch normalization layer placement order based on asymmetric convolution optimization in accordance with the present invention.
Detailed Description
In order that the invention may be more readily understood, a more particular description of the invention will be rendered by reference to specific embodiments that are illustrated in the appended drawings.
As shown in fig. 1 and 2, a convolutional neural network training method based on an asymmetric convolutional kernel includes the following steps:
s01: acquiring a data set, wherein the data set is used for training a convolutional neural network;
s02: splitting n-n convolution kernels of a designated convolution layer in a convolution neural network into n-1 and 1*n asymmetric convolution kernels to form two parallel convolution layers, and adding a batch normalization layer after the split two parallel convolution layers;
s03: training the convolutional neural network processed in the step S02 by adopting the data set acquired in the step S01, acquiring an image feature map in the training process, extracting image features, and classifying images;
s04: after the convolutional neural network is trained in one round, the convolutional neural network is verified by using a verification set;
s05: and (3) circularly executing the steps S2 to S4, and performing multi-round training on the convolutional neural network until reaching the ending condition of model convergence of the convolutional neural network.
Specifically, in step S01, data enhancement is performed on the data set to obtain a larger-scale data set for application to neural network model testing.
Specifically, the data enhancement mode is as follows: at least one of blurring and rotation.
Specifically, in step S2, after splitting the n×n convolution kernels of the specified convolution layer into two asymmetric convolution kernels, a padding operation is added to ensure that the size of the convolution kernels is consistent with that of the normal convolution output.
Specifically, the filling formula of the filling operation is:
wherein W and H are pattern width and pattern height;
KW, KH is the convolution kernel width and convolution kernel height;
PW, PH is left and right filling size and up and down filling size;
s is convolution step length; w N,HN is the updated pattern width and pattern height.
In the asymmetric convolution, padding is only required in the vertical or horizontal direction, and padding is not required to be applied in the other direction, so that the size of the output feature map and the size of the input feature map can be ensured to be consistent.
The batch normalization layer (Batch Normalization) pulls back the distribution of any neuron of each layer of neural network, namely input value, to the standard normal distribution with the mean value of 0 and the variance of 1 through a certain normalization means, and forcibly pulls back the distribution with more and more deviation to the distribution of the comparison standard, so that the activated input value falls in a region with more sensitive nonlinear function to input, the gradient is increased, the problem of gradient disappearance is avoided, the gradient is increased, the learning convergence speed is high, and the training speed is accelerated.
Specifically, the BN formula for the batch normalization layer is:
yi＝BNγ,β(xi)；
wherein,
I is an index number in the batch data;
x i is the ith sample input data in the batch data;
y i is the i-th sample output data;
Gamma and beta are self-learning parameters, scaling factors and offset factors, offset is carried out on the scaled data, and the positions of the data are adjusted so that the model learns more complex characteristic representation;
m is the total number of index batch input samples for a sample in a batch of input samples.
Mu Β is the mean of the batch input data;
variance of input data for a batch;
For normalized data, the distribution range and mean variance are adjusted to learn the model to the appropriate feature representation. ;
E is a very small number to avoid dividing by 0.
The batch normalization layer is arranged after the split convolution layer and is not suitable for being arranged in the middle of the split parallel convolution layer; if placed in the middle of the split parallel convolution layers, the batch normalization layer makes the model less sensitive to the initialization method and parameters in the network, but conflicts with the perceptibility of the directional information in the input data.
Specifically, the method further comprises the steps of:
S06: calculating the accuracy and positive and negative sample prediction bias values of the convolutional neural network on the test set, wherein the convolutional neural network uses the accuracy and the positive and negative sample prediction bias values together as evaluation indexes, and the final algorithm target of the convolutional neural network is as follows: the predicted deviation value of the positive and negative samples is the lowest; the positive sample prediction deviation and the negative sample prediction deviation can be calculated according to the probability of predicting each category by the model of the convolutional neural network.
Specifically, in step S6, when the classification accuracy is similar, performance evaluation is performed in combination with the positive and negative sample prediction bias.
Wherein,
In the formula, out is the probability of model prediction category; i is that the model predicts that the current image belongs to the ith category; out i is the probability that the model predicts that the current image belongs to the ith category; max is the subscript of the maximum value in Out, i.e. the category to which the model prediction belongs; cur is the true class of the current image.
Specifically, in this embodiment, in step S3, in the forward propagation process of the convolutional neural network, the calculation formula of each layer of input features:
Wherein y (l) is the output of the first convolutional layer;
Where x is the input vector;
wherein, Is convolution operation;
Wherein b l is offset;
Wherein w is the weight value of the layer;
Wherein m is a set of input features;
Wherein i is the ith neuron;
Where f (x) is the activation function.
Epochs was set to 100, batchSize to 128, and LR to 0.0009 in ResNet model based on asymmetric convolution optimization. Wherein tables 1-3 are the correct rate of the model test set
Table 1 performance of ResNet model based on asymmetric convolution optimization on the accuracy of the Raw-img test set:
Table 2 behavior of ResNet model based on asymmetric convolution optimization on positive and negative prediction bias in the Raw-img test set:
Table 3 behavior of ResNet model based on asymmetric convolution optimization for test set correctness and positive and negative prediction bias at IMAGENETTE:
the technical problems, technical solutions and advantageous effects solved by the present invention have been further described in detail in the above-described embodiments, and it should be understood that the above-described embodiments are only illustrative of the present invention and are not intended to limit the present invention, and any modifications, equivalent substitutions, improvements, etc. within the spirit and principle of the present invention should be included in the scope of protection of the present invention.
Claims (7)
1. A convolutional neural network training method based on an asymmetric convolutional kernel is characterized by comprising the following steps:
s01: acquiring a data set, wherein the data set is used for training a convolutional neural network;
s02: splitting n-n convolution kernels of a designated convolution layer in a convolution neural network into n-1 and 1*n asymmetric convolution kernels to form two parallel convolution layers, and adding a batch normalization layer after the split two parallel convolution layers;
s03: training the convolutional neural network processed in the step S02 by adopting the data set acquired in the step S01, acquiring an image feature map in the training process, extracting image features, and classifying images;
s04: after the convolutional neural network is trained in one round, the convolutional neural network is verified by using a verification set;
s05: and (3) circularly executing the steps S2 to S4, and performing multi-round training on the convolutional neural network until reaching the ending condition of model convergence of the convolutional neural network.
2. The convolutional neural network training method of claim 1, wherein in step S01, data enhancement is performed on the data set.
3. The convolutional neural network training method of claim 2,
The data enhancement mode is as follows: at least one of blurring and rotation.
4. The convolutional neural network training method of claim 1,
In step S2, after splitting the n×n convolution kernels of the specified convolution layer into two asymmetric convolution kernels, a padding operation is added.
5. The convolutional neural network training method of claim 4,
The filling formula of the filling operation is:
wherein,
W, H is the width and height of the pattern;
KW, KH is the convolution kernel width and convolution kernel height;
PW, PH is left and right filling size and up and down filling size;
s is convolution step length; w N,HN is the updated pattern width and pattern height.
6. The convolutional neural network training method of claim 1,
The BN formula for the batch normalization layer is:
yi＝BNγ,β(xi)；
wherein,
I is an index number in the batch data;
x i is the ith sample input data in the batch data;
y i is the i-th sample output data;
gamma, beta are self-learning parameters;
m is the total number of index batch input samples for a sample in a batch of input samples.
Mu Β is the mean of the batch input data;
variance of input data for a batch;
Is normalized data;
E is a very small number to avoid dividing by 0.
7. The convolutional neural network training method of claim 1, further comprising:
S06: calculating the accuracy and positive and negative sample prediction bias values of the convolutional neural network on the test set, wherein the convolutional neural network uses the accuracy and the positive and negative sample prediction bias values together as evaluation indexes, and the final algorithm target of the convolutional neural network is as follows: the positive and negative samples have the lowest predicted bias value.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410383952.XA CN118114723A (en) | 2024-04-01 | 2024-04-01 | Convolutional neural network training method based on asymmetric convolution kernel |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410383952.XA CN118114723A (en) | 2024-04-01 | 2024-04-01 | Convolutional neural network training method based on asymmetric convolution kernel |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN118114723A true CN118114723A (en) | 2024-05-31 |
Family
ID=91214199
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410383952.XA Pending CN118114723A (en) | 2024-04-01 | 2024-04-01 | Convolutional neural network training method based on asymmetric convolution kernel |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN118114723A (en) |
-
2024
- 2024-04-01 CN CN202410383952.XA patent/CN118114723A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108564129B (en) | Trajectory data classification method based on generation countermeasure network | |
| US11804074B2 (en) | Method for recognizing facial expressions based on adversarial elimination | |
| CN108615010B (en) | Facial expression recognition method based on parallel convolution neural network feature map fusion | |
| CN111783782B (en) | Remote sensing image semantic segmentation method fusing and improving UNet and SegNet | |
| CN103605972B (en) | Non-restricted environment face verification method based on block depth neural network | |
| CN112116605A (en) | Pancreas CT image segmentation method based on integrated depth convolution neural network | |
| CN109063719B (en) | Image classification method combining structure similarity and class information | |
| CN112464865A (en) | Facial expression recognition method based on pixel and geometric mixed features | |
| CN108960301B (en) | Ancient Yi-nationality character recognition method based on convolutional neural network | |
| CN112418397B (en) | Image classification method based on lightweight convolutional neural network | |
| CN108052959A (en) | A kind of method for improving deep learning picture recognition algorithm robustness | |
| CN113255832B (en) | Method for identifying long tail distribution of double-branch multi-center | |
| CN115035418A (en) | Remote sensing image semantic segmentation method and system based on improved deep LabV3+ network | |
| CN112232395B (en) | Semi-supervised image classification method for generating countermeasure network based on joint training | |
| CN106503616A (en) | A kind of Mental imagery Method of EEG signals classification of the learning machine that transfinited based on layering | |
| CN113228062A (en) | Deep integration model training method based on feature diversity learning | |
| CN109101984B (en) | Image identification method and device based on convolutional neural network | |
| CN111144296B (en) | Retina fundus picture classification method based on improved CNN model | |
| CN113128455A (en) | Cell image reconstruction model training method and system | |
| CN114241564A (en) | Facial expression recognition method based on inter-class difference strengthening network | |
| CN114821736A (en) | Multi-modal face recognition method, device, equipment and medium based on contrast learning | |
| CN114120401A (en) | Human face anti-fraud method based on cross-domain feature alignment network | |
| CN117315534A (en) | Short video classification method based on VGG-16 and whale optimization algorithm | |
| CN112836729A (en) | Construction method of image classification model and image classification method | |
| CN115830401B (en) | Small sample image classification method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |