CN111160474B - Image recognition method based on deep course learning - Google Patents
Image recognition method based on deep course learning Download PDFInfo
- Publication number
- CN111160474B CN111160474B CN201911401914.8A CN201911401914A CN111160474B CN 111160474 B CN111160474 B CN 111160474B CN 201911401914 A CN201911401914 A CN 201911401914A CN 111160474 B CN111160474 B CN 111160474B
- Authority
- CN
- China
- Prior art keywords
- network
- training
- teacher
- student
- classification
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an image recognition method based on deep course learning, which belongs to the field of image recognition and comprises the following steps: constructing teacher and student networks based on the deep convolutional neural network; training the teacher network by using the training samples to carry out image classification training, and predicting the probability that the training samples belong to each category; calculating the difference between the prediction and the label of the teacher network to update the parameters; transmitting the forecast information to a student network; training a student network; guiding the student network training by using the prediction information result of the teacher network; calculating a difference update parameter between a student network prediction result and a label; completing student network classification training; the training student network realizes the identification and classification of the images. The invention simulates the process of easy-to-get-difficult learning of human beings, has reasonable training process, greatly reduced workload and quick network parameter updating, and has higher prediction precision and more reliable and stable performance due to the influence of gradient difference unbalance samples generated by different samples.
Description
Technical Field
The invention relates to the technical field of image recognition, in particular to an image recognition method based on deep course learning.
Background
Due to the large increase in computational power and the increase in training data, the training conditions of the deep neural network are satisfied, and due to the strong fitting capability, the deep neural network achieves the best effect in many tasks. However, since parameters of the deep neural network are updated based on a large number of samples, the samples play a critical role in the training process. Therefore, the training work intensity is high, the training process is complex, and the parameter updating of the deep neural network is slow.
Meanwhile, the deep convolutional neural network calculates the gradient of each parameter according to the prediction result of the training sample and the correctly classified labels by the deep convolutional neural network, and then the generated gradient is used for updating the network parameters. In the prior art, the difference between the predicted result and the correct labeling is large, so that the deep convolutional neural network is punished, and the true labeling result cannot be fitted.
Therefore, an image recognition method which is simple in training process, fast in parameter updating of the deep neural network and capable of fitting the real labels of the images is urgently needed.
Disclosure of Invention
The invention aims to provide an image recognition method which has simple training process, fast parameter updating of a deep neural network and capability of fitting the real annotation of an image, and the invention comprises the following steps:
an image recognition method based on deep course learning comprises the following steps:
s10, constructing a teacher network and a student network based on a deep convolutional neural network;
s20, performing image recognition and classification training on the teacher network by using a training sample; the teacher network learns the training samples and predicts the probability that the training samples belong to each category;
s30, calculating the difference between the prediction information result of the teacher network and the training sample label by using a cross entropy loss function, and updating the parameters of the teacher network;
s40, transmitting a predicted information result of the training sample to the student network by the teacher network;
s50, performing image classification training on the student network by using the same training sample;
s60, guiding the student network to train by taking a predicted information result of the teacher network on the training sample as a judging basis of difficulty level;
s70, calculating the difference between the student network classification prediction result and the training sample label by using a BLCL loss function, and updating parameters of the student network;
s80, completing the student network classification training;
and S90, based on the training-completed student network, the identification and classification of the images are realized.
Further, the step S20 of predicting the probability of belonging to each category by the teacher network is: the deep convolutional neural network of the teacher network processes the pictures of the training samples into visual features, generates scoring of corresponding classifications through global pooling and a full-connection network, synthesizes the scoring of different classifications by using a softmax function, converts the scoring into the probability of classification, and predicts the probability that the pictures belong to each classification; the softmax function is
Wherein x represents the classifying target picture, Z k A score representing the picture being classified into the kth category.
Further, the step S30 specifically includes: after the deep convolutional neural network is classified, calculating the difference between the prediction probability and the real label by using a cross entropy loss function, and calculating the gradient size and the gradient direction which are needed to be updated for each parameter in the deep convolutional neural network from the difference by using a back propagation algorithm; calculating the loss of a whole batch of pictures and the corresponding gradient each time by adopting a batch gradient random descent algorithm, then updating the loss of the whole batch of pictures and the corresponding gradient, and updating parameters; the loss function is
Where N is the number of samples of the batch data, x i To attempt to classify the sample, y i For the corresponding label, θ t For the teacher's network to be in the training process,is a cross entropy loss function.
Further, the BLCL loss function in step S70 is
Wherein x is i In order to train the sample,the prediction result of the teacher network is that the student network judges the difficulty of the sample by using the prediction information of the teacher network and carries out adaptive weight on the sample, and L is a cross entropy loss function or an L2 loss function.
The invention has the beneficial effects that:
1. the invention simulates the process of easy-to-get-difficult human learning, firstly learns easy-to-classify samples, then learns difficult samples, has reasonable training process and greatly reduces training workload.
2. The training of the deep convolutional neural network is based on a random gradient descent method, and in each training round, gradients are calculated according to loss generated by the prediction result and the distance of the correct label, and then the generated gradients are used for updating network parameters, so that the network parameter updating speed is high.
3. In the training process of the deep neural convolutional network, loss and gradient sizes generated by different samples are monitored in real time, and then the influence of the samples is balanced according to gradient differences generated by the different samples. Compared with the original unbalanced training mode, the method has higher prediction precision and more reliable and stable performance.
Drawings
FIG. 1 is a flow chart of an image recognition method based on deep course learning
Detailed Description
In this embodiment, given an image dataset such as CIFAR10, CIFAR100 uses different deep neural network models to verify the effectiveness of BLCL learning.
Each data set is divided into a training set and a testing set, and training data is trained in batches.
The method comprises the following specific steps:
s10, constructing a teacher network and a student network based on a deep convolutional neural network; convolutional neural network this embodiment employs ResNet-8, resNet-20, resNet-32 and DensNet-BC.
S20, performing image recognition and classification training on a teacher network by using training samples; the teacher network learns the training samples and predicts the probability that the training samples belong to each category;
the training sample is CIFAR10 and CIFAR100 data set, the training set comprises 50,000 pictures, the test set comprises 10,000 pictures, and each training picture is subjected to data disturbance enhancement.
The teacher network predicts the probability of belonging to each category as follows: the deep convolutional neural network of the teacher network processes the pictures of the training samples into visual features, generates scoring corresponding to the classification through global pooling and a fully connected network, synthesizes the scoring of different classes by using a softmax function, converts the scoring into the probability of classification, and predicts the probability that the pictures belong to each class; the softmax function is
Wherein x represents the classifying target picture, Z k Scoring representing the pictures being classified into a kth category
S30, calculating the difference between the prediction information result of the teacher network and the training sample label by using the cross entropy loss function, and updating the parameters of the teacher network;
after the deep convolutional neural network is classified, calculating the difference between the prediction probability and the real label by using a cross entropy loss function, and calculating the gradient size and the gradient direction which are needed to be updated for each parameter in the deep convolutional neural network from the difference by using a back propagation algorithm; in order to increase the efficiency and stability of network parameter updating, a batch gradient random descent algorithm is adopted, the loss and the corresponding gradient of a whole batch of pictures are calculated each time, then the loss and the corresponding gradient of the whole batch of pictures in updating are carried out, and parameter updating is carried out; the loss function is
Where N is the number of samples of the batch data, x i To attempt to classify the sample, y i For the corresponding label, θ t For the teacher's network to be in the training process,is a cross entropy loss function;
the batch size of each batch of data was 128, optimized by random gradient descent.
S40, transmitting a predicted information result of the training sample to a student network by the teacher network;
s50, performing image classification training on the student network by using the same training sample;
s60, guiding the student to train by taking the prediction information result of the teacher network on the training sample as a judgment basis of the difficulty level;
s70, calculating the difference between the classification prediction result of the student network and the training sample label by using the BLCL loss function, and updating parameters of the student network;
BLCL loss function of
Wherein x is i In order to train the sample,the prediction result of the teacher network is that the student network judges the difficulty of the sample by using the prediction information of the teacher network and carries out adaptive weight on the sample, and L is a cross entropy loss function or an L2 loss function.
S80, finishing student network classification training;
and S90, based on the training-completed student network, the identification and classification of the images are realized.
The implementation process of the embodiment is as follows: firstly, training set data are sent into a teacher network model, the difference between a prediction result and a label is calculated by using a cross entropy loss function, and parameters of the teacher network model are updated. And then, taking the result of the teacher network model prediction as a judgment basis of the sample difficulty level to guide the training of the student network model. And the same batch of training data is sent into the student network model to calculate the difference between the prediction result and the label according to the BLCL loss function to update the parameters of the student model according to the judgment of the teacher network model, so that the weight of the sample easy to classify is increased, and the weight of the sample difficult to classify is reduced. And the student classification model is used as a training result, and finally the student classification network model can realize image recognition.
The foregoing description is only a preferred embodiment of the present invention, and is not intended to limit the technical scope of the present invention, so any minor modifications, equivalent changes and modifications made to the above embodiments according to the technical principles of the present invention still fall within the scope of the technical solutions of the present invention.
Claims (2)
1. An image recognition method based on deep course learning is characterized by comprising the following steps:
s10, constructing a teacher network and a student network based on a deep convolutional neural network;
s20, performing image recognition and classification training on the teacher network by using a training sample; the teacher network learns the training samples and predicts the probability that the training samples belong to each category;
s30, calculating the difference between the prediction information result of the teacher network and the training sample label by using a cross entropy loss function, and updating the parameters of the teacher network;
s40, transmitting a predicted information result of the training sample to the student network by the teacher network;
s50, performing image classification training on the student network by using the same training sample;
s60, guiding the student network to train by taking a predicted information result of the teacher network on the training sample as a judging basis of difficulty level;
s70, calculating the difference between the student network classification prediction result and the training sample label by using a BLCL loss function, and updating parameters of the student network;
s80, completing the student network classification training;
s90, based on the training-completed student network, identifying and classifying images;
the step S20 of predicting the probability of belonging to each category by the teacher network includes: the deep convolutional neural network of the teacher network processes the pictures of the training samples into visual features, generates scoring of corresponding classifications through global pooling and a full-connection network, synthesizes the scoring of different classifications by using a softmax function, converts the scoring into the probability of classification, and predicts the probability that the pictures belong to each classification; the softmax function is
Wherein x represents the classifying target picture, Z k A score representing the picture being classified into a kth category;
the training samples are CIFAR10 and CIFAR100 data sets, wherein the training set comprises 50,000 pictures, the test set comprises 10,000 pictures, and each training picture is subjected to data disturbance enhancement;
the BLCL loss function in step S70 is
Wherein x is i In order to train the sample,the prediction result of the teacher network is that the student network judges the difficulty of the sample by using the prediction information of the teacher network and carries out adaptive weight on the sample, and L is a cross entropy loss function or an L2 loss function.
2. The image recognition method according to claim 1, wherein the step S30 specifically comprises the steps of: after the deep convolutional neural network is classified, calculating the difference between the prediction probability and the real label by using a cross entropy loss function, and calculating the gradient size and the gradient direction which are needed to be updated for each parameter in the deep convolutional neural network from the difference by using a back propagation algorithm; calculating the loss of a whole batch of pictures and the corresponding gradient each time by adopting a batch gradient random descent algorithm, then updating the loss of the whole batch of pictures and the corresponding gradient, and updating parameters; the loss function is
Where N is the number of samples of the batch data, x i To attempt to classify the sample, y i For the corresponding label, θ t For the teacher's network to be in the training process,is a cross entropy loss function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911401914.8A CN111160474B (en) | 2019-12-30 | 2019-12-30 | Image recognition method based on deep course learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911401914.8A CN111160474B (en) | 2019-12-30 | 2019-12-30 | Image recognition method based on deep course learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111160474A CN111160474A (en) | 2020-05-15 |
| CN111160474B true CN111160474B (en) | 2023-08-29 |
Family
ID=70559311
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911401914.8A Active CN111160474B (en) | 2019-12-30 | 2019-12-30 | Image recognition method based on deep course learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111160474B (en) |
Families Citing this family (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111639710B (en) * | 2020-05-29 | 2023-08-08 | 北京百度网讯科技有限公司 | Image recognition model training method, device, equipment and storage medium |
| CN111709480B (en) * | 2020-06-17 | 2023-06-23 | 北京百度网讯科技有限公司 | Method and device for identifying image category |
| CN111814689A (en) * | 2020-07-09 | 2020-10-23 | 浙江大华技术股份有限公司 | Fire recognition network model training method, fire recognition method and related equipment |
| CN111814963B (en) * | 2020-07-17 | 2024-05-07 | 中国科学院微电子研究所 | Image recognition method based on deep neural network model parameter modulation |
| CN112163450A (en) * | 2020-08-24 | 2021-01-01 | 中国海洋大学 | Based on S3High-frequency ground wave radar ship target detection method based on D learning algorithm |
| CN112308130B (en) * | 2020-10-29 | 2021-10-15 | 成都千嘉科技有限公司 | Deployment method of deep learning network of Internet of things |
| CN112381209B (en) * | 2020-11-13 | 2023-12-22 | 平安科技（深圳）有限公司 | Model compression method, system, terminal and storage medium |
| CN112418343B (en) * | 2020-12-08 | 2024-01-05 | 中山大学 | Multi-teacher self-adaptive combined student model training method |
| CN112560948B (en) * | 2020-12-15 | 2024-04-26 | 中南大学 | Fundus image classification method and imaging method under data deviation |
| CN112508943A (en) * | 2020-12-25 | 2021-03-16 | 四川工商学院 | Breast tumor identification method based on ultrasonic image |
| CN115019079B (en) * | 2021-03-04 | 2024-05-28 | 北京大学 | Method for accelerating deep learning training by distributed outline optimization for image recognition |
| CN112990298B (en) * | 2021-03-11 | 2023-11-24 | 北京中科虹霸科技有限公司 | Key point detection model training method, key point detection method and device |
| CN113111968B (en) * | 2021-04-30 | 2024-03-22 | 北京大米科技有限公司 | Image recognition model training method, device, electronic equipment and readable storage medium |
| CN113326764B (en) * | 2021-05-27 | 2022-06-07 | 北京百度网讯科技有限公司 | Method and device for training image recognition model and image recognition |
| CN113284122B (en) * | 2021-05-31 | 2023-06-23 | 五邑大学 | Roll paper packaging defect detection method and device based on deep learning and storage medium |
| CN113094758B (en) * | 2021-06-08 | 2021-08-13 | 华中科技大学 | Gradient disturbance-based federated learning data privacy protection method and system |
| CN113222123A (en) * | 2021-06-15 | 2021-08-06 | 深圳市商汤科技有限公司 | Model training method, device, equipment and computer storage medium |
| CN113792574B (en) * | 2021-07-14 | 2023-12-19 | 哈尔滨工程大学 | Cross-dataset expression recognition method based on metric learning and teacher student model |
| CN113610118B (en) * | 2021-07-19 | 2023-12-12 | 中南大学 | Glaucoma diagnosis method, device, equipment and method based on multitasking course learning |
| CN113283848B (en) * | 2021-07-21 | 2021-09-28 | 湖北浩蓝智造科技有限公司 | Goods warehousing detection method, warehousing system and storage medium |
| CN113688760A (en) * | 2021-08-31 | 2021-11-23 | 广州文远知行科技有限公司 | Automatic driving data identification method and device, computer equipment and storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109034205A (en) * | 2018-06-29 | 2018-12-18 | 西安交通大学 | Image classification method based on the semi-supervised deep learning of direct-push |
| CN110163344A (en) * | 2019-04-26 | 2019-08-23 | 北京迈格威科技有限公司 | Neural network training method, device, equipment and storage medium |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180268292A1 (en) * | 2017-03-17 | 2018-09-20 | Nec Laboratories America, Inc. | Learning efficient object detection models with knowledge distillation |
| CN107247989B (en) * | 2017-06-15 | 2020-11-24 | 北京图森智途科技有限公司 | Real-time computer vision processing method and device |
-
2019
- 2019-12-30 CN CN201911401914.8A patent/CN111160474B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109034205A (en) * | 2018-06-29 | 2018-12-18 | 西安交通大学 | Image classification method based on the semi-supervised deep learning of direct-push |
| CN110163344A (en) * | 2019-04-26 | 2019-08-23 | 北京迈格威科技有限公司 | Neural network training method, device, equipment and storage medium |
Non-Patent Citations (1)
| Title |
|---|
| 基于DCGAN反馈的深度差分隐私保护方法;毛典辉;李子沁;蔡强;薛子育;;北京工业大学学报(06);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111160474A (en) | 2020-05-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111160474B (en) | Image recognition method based on deep course learning | |
| CN114241282B (en) | Knowledge distillation-based edge equipment scene recognition method and device | |
| CN110598029B (en) | Fine-grained image classification method based on attention transfer mechanism | |
| WO2020046213A1 (en) | A method and apparatus for training a neural network to identify cracks | |
| CN108416370A (en) | Image classification method, device based on semi-supervised deep learning and storage medium | |
| CN111598164B (en) | Method, device, electronic equipment and storage medium for identifying attribute of target object | |
| JP2020123330A (en) | Method for acquiring sample image for label acceptance inspection from among auto-labeled images utilized for neural network learning, and sample image acquisition device utilizing the same | |
| CN112685504B (en) | Production process-oriented distributed migration chart learning method | |
| CN112905997B (en) | Method, device and system for detecting poisoning attack facing deep learning model | |
| CN107247952B (en) | Deep supervision-based visual saliency detection method for cyclic convolution neural network | |
| CN112132014A (en) | Target re-identification method and system based on non-supervised pyramid similarity learning | |
| CN113326826A (en) | Network model training method and device, electronic equipment and storage medium | |
| CN114782752B (en) | Small sample image integrated classification method and device based on self-training | |
| CN110210695A (en) | A kind of tower control simulated training appraisal procedure based on support vector machines | |
| CN115187772A (en) | Training method, device and equipment of target detection network and target detection method, device and equipment | |
| CN114170484B (en) | Picture attribute prediction method and device, electronic equipment and storage medium | |
| CN117690164B (en) | Airport bird identification and driving method and system based on edge calculation | |
| CN115063664A (en) | Model learning method, training method and system for industrial vision detection | |
| CN113837268B (en) | Method, device, equipment and medium for determining track point state | |
| CN112329830B (en) | Passive positioning track data identification method and system based on convolutional neural network and transfer learning | |
| CN111797935B (en) | Semi-supervised depth network picture classification method based on group intelligence | |
| CN111191027B (en) | Generalized zero sample identification method based on Gaussian mixture distribution (VAE) | |
| CN112926052A (en) | Deep learning model security vulnerability testing and repairing method, device and system based on genetic algorithm | |
| Shao et al. | Attention GAN-based method for designing intelligent making system | |
| CN115858388A (en) | Test case priority ordering method and device based on variation model mapping chart |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |