CN108647702B - Large-scale food material image classification method based on transfer learning - Google Patents

Large-scale food material image classification method based on transfer learning Download PDF

Info

Publication number
CN108647702B
CN108647702B CN201810332217.0A CN201810332217A CN108647702B CN 108647702 B CN108647702 B CN 108647702B CN 201810332217 A CN201810332217 A CN 201810332217A CN 108647702 B CN108647702 B CN 108647702B
Authority
CN
China
Prior art keywords
tree
food material
model
image
categories
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201810332217.0A
Other languages
Chinese (zh)
Other versions
CN108647702A (en
Inventor
肖光意
刘欢
刘毅
吴淇
黄宗杰
陈浩
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hunan University
Original Assignee
Hunan University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hunan University filed Critical Hunan University
Priority to CN201810332217.0A priority Critical patent/CN108647702B/en
Publication of CN108647702A publication Critical patent/CN108647702A/en
Application granted granted Critical
Publication of CN108647702B publication Critical patent/CN108647702B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/24323—Tree-organised classifiers
    • 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)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

本发明公开了一种基于迁移学习的大规模食材图像分类方法，包括以下步骤：输入一张原始图片，通过迁移学习的知识，输出包含两方面的信息，其一是对收货环境的识别结果，其二是对食材种类的识别结果；在基于多任务的Prior‑Tree CNN模型的前向传播之后，由最后一个全连接层产生的联合特征将被输入到不同的分类任务中；本发明的基于迁移学习的大规模食材图像分类方法，基于迁移学习方法对卷积神经网络模型进行改进，并且提出Prior‑Tree CNN模型来学习类结构和分类器参数，改进后的模型具有高效性、高准确率及普适性。

Figure 201810332217

The invention discloses a large-scale food image classification method based on migration learning, comprising the following steps: inputting an original image, and outputting information including two aspects through the knowledge of migration learning, one is the recognition result of the receiving environment , and the second is the recognition result of the type of food; after the forward propagation of the multi-task-based Prior-Tree CNN model, the joint features generated by the last fully connected layer will be input into different classification tasks; A large-scale food image classification method based on transfer learning, the convolutional neural network model is improved based on the transfer learning method, and the Prior-Tree CNN model is proposed to learn the class structure and classifier parameters. The improved model has high efficiency and high accuracy rate and universality.

Figure 201810332217

Description

Large-scale food material image classification method based on transfer learning
Technical Field
The invention relates to a large-scale food material image classification method based on transfer learning, and belongs to the technical field of food material image classification.
Background
With the rapid development of deep learning, the image classification technology has developed into a more popular research direction in the computer vision field at home and abroad, and is applied to more and more fields; based on an actual scene, how to construct a method for classifying images, and the method with high accuracy and good robustness is an unavoidable core problem in the image classification technology moving to practical application.
The training classifier needs a large number of samples, when the training samples are few, the problem is solved, and for the serious problem of unbalanced category, the inherent structure of a group of classes in the integrity environment data set is used for presenting a priority tree CNN model; for example, four food materials need to be purchased, large red pepper is related to the screw pepper, the small rice pepper and the sharp green pepper, but only one image of the screw pepper is needed, which makes it difficult to train one classifier to distinguish hundreds of pictures of the screw pepper and other three categories; by the marking examples of the related categories, the model can learn a task from one screw pepper example more easily through zero-shot; learning new categories by transferring knowledge from relevant categories, and only knowing the characteristics of the Zanthoxylum piperitum; the traditional image classification comprises methods such as KNN and SVM, the KNN is classified by measuring distances among different characteristic values, and the defects of the KNN comprise large calculated amount, large prediction deviation when a sample is unbalanced and the like; the SVM belongs to supervised learning, is a two-classification model, the basic model of the SVM is a linear classifier with the maximum interval defined on a feature space, and the defects of the SVM comprise that the interpretation force of high-dimensional mapping of a kernel function is not strong, particularly a radial basis function; only a two-class classification algorithm is given, so that the problem of multi-classification is difficult to solve; aiming at the problems in the Prior art, the convolutional neural network model (CNN) is improved based on a migration learning method, a Prior-Tree CNN model is provided for learning class structures and classifier parameters, a large amount of existing labeled data of related classes are fully utilized, and efficient knowledge migration of the related classes is learned, so that the accuracy and robustness of object classification of a small sample set are improved by using a classifier obtained by training on a large sample set, and a Meal-53 small sample data set is established, and the aim of improving the classification accuracy is fulfilled under the condition that only a few samples exist.
Disclosure of Invention
In order to solve the problems, the invention provides a large-scale food material image classification method based on transfer learning, a convolutional neural network model (CNN) is improved based on the transfer learning method, a primer-Tree CNN model is provided for learning class structures and classifier parameters, and the improved model has high efficiency, high accuracy and universality.
The invention discloses a large-scale food material image classification method based on transfer learning, which comprises the following steps of: inputting an original picture, and outputting information containing two aspects through the transfer learning knowledge, wherein one is the identification result of the receiving environment, and the other is the identification result of the food material type; after forward propagation of the multitasking based Prior-Tree CNN model, the joint features generated by the last fully connected layer will be input into different classification tasks.
Further, the classification method specifically comprises the following steps: adding an artificially set tree structure on the multi-task classification CNN model, firstly setting a parent node according to the tree structure, then enabling a newly added subclass to directly obtain the characteristics of the parent, and training under the condition; therefore, even if the data is less, the newly added subclasses can have good effects; the model can also be viewed as a specific softmax penalty for different tasks; after optimizing the joint loss layer and sharing the visual features, the CNN model will deliver its relevant parameters in the back propagation process; marking whether the collected food material images are pictures in a clean goods receiving environment by using c e to {0, 1}, and marking the food material types by using a multi-type label K to e to { 1.. K }, wherein K is the number of all the types; finally, training the whole network and parameters iteratively until convergence;
in the actual process, the learning process is almost the same as the standard gradient reduction, and during testing, whether the image is a clean receiving environment picture is verified firstly; if the image is predicted to be a non-clean image, the system will take a new image to continue the process until a clean image is found within a given time; second, if the image is predicted to be a vegetable category, the model will filter other labels in the vegetable in the given order; finally, through the Priori Tree CNN model combined with the information of orders, weight and the like, the clean image can obtain the food material category and the predicted score.
Still further, the formula expression of the color-Tree CNN model is as follows:
firstly, supposing that a prior tree is provided, all food categories are composed of three layers of trees, the prior tree has K +1 leaf nodes, and K food category labels and 1 neglected label correspond to the prior tree respectively; the leaf nodes are connected with 7 father nodes of the second layer, the K food material categories have 4 fathers and comprise S1, S2, S3 and S4, and the receiving environment categories are divided into 3 groups comprising N1, N2 and N3; for the classification, if the input images are classified into vegetables with high confidence level, other categories in the order are filtered for prediction, the categories such as meat, aquatic products and the like do not need to be concerned, and the accuracy rate of the food material category identification task is further improved;
in order to simplify the Priori Tree model, only the relationships between the labels in the K class and the four parent nodes thereof are concerned; before the softmax function, the leaf tag node k is associated with a weight vector
Figure BDA0001628267910000031
Associating; each super class node s and vector
Figure BDA0001628267910000032
Associating, wherein s belongs to {1, 2, 3, 4 }; e.g. betacabbageAnd betacarrotRecord and alphavegatableThe deviation therebetween; the following generative model is defined for β:
Figure BDA0001628267910000033
this prior expression describes the relationship between classes, with the conditional distribution over k being expressed as:
Figure BDA0001628267910000034
inferring the value of { W, β, α } from MAP, maximizing:
logP(k|I，t，W，β)+logP(k|I，W，β)+logP(W)+logP(β|α)+logP(α)； (3)
from the point of view of the loss function, minimize:
Figure BDA0001628267910000041
here, by fixing the value of α to 0, the loss function will be reduced to the standard loss function, let
CsThen { k | paramt (k) ═ s }, then
Figure BDA0001628267910000042
Therefore, the loss function in equation (4) is optimized by iteratively performing the following two steps; first, W and β are maximized, α is fixed using a standard random gradient descent (SGD) of a standard loss function; second, β is fixed by equation (5), thereby maximizing α.
Compared with the Prior art, the migration learning-based large-scale food material image classification method improves the convolutional neural network model based on the migration learning method, and provides the Prior-Tree CNN model to learn the class structure and the classifier parameters, and the improved model has the following advantages: the method has high efficiency, provides a multitask CNN model to classify honest/dishonest environment images so as to reduce noise in a Meal-53 data set and more effectively utilize the characteristics of training samples; the accuracy is high, and the improved model can correctly identify the category when the data training set is less; generally, the improved CNN model solves the problem through learning, is not limited to a specific problem, and can automatically build a model according to the problem, thereby solving similar problems.
Drawings
FIG. 1 is a schematic diagram of the overall architecture of the Prior-Tree CNN model of the present invention; wherein, the diagram (a) is a Prior-Tree CNN model diagram; fig. b is a tree structure diagram.
FIG. 2 is a schematic diagram of the filtering of tags within a model by super class prediction of the present invention; wherein, the diagram (a) is a priori tree diagram; FIG. b is a schematic view of order classification.
FIG. 3 is a schematic illustration of the accuracy of validation using 1-crop on the Mealcome dataset of the present invention.
Figure 4 is a graphical representation of the accuracy of validation using 10-crops on the Mealcome dataset of the present invention.
FIG. 5 is a schematic diagram of an exemplary structure of a Priori Tree Model prediction sample of the present invention.
Detailed Description
The invention discloses a large-scale food material image classification method based on transfer learning, which comprises the following steps of: in order to enhance the effect of the model on classification and enable the model to correctly identify the class when the training set of data is less; a primitive picture Ii is input, and information containing two aspects is output through the knowledge of transfer learning, wherein one is an identification result of a receiving environment, and the other is an identification result of a food material type; after forward propagation of the multitasking based Prior-Tree CNN model, the joint features generated by the last fully connected layer will be input into different classification tasks.
The classification method specifically comprises the following steps: adding an artificially set tree structure on the multi-task classification CNN model, firstly setting a parent node according to the tree structure, then enabling a newly added subclass to directly obtain the characteristics of the parent, and training under the condition; therefore, even if the data is less, the newly added subclasses can have good effects; the model can also be viewed as a specific softmax penalty for different tasks; after optimizing the joint loss layer and sharing the visual features, the CNN model will deliver its relevant parameters in the back propagation process; marking whether the collected food material images are pictures in a clean goods receiving environment by using c e to {0, 1}, and marking the food material types by using a multi-type label K to e to { 1.. K }, wherein K is the number of all the types; finally, training the whole network and parameters iteratively until convergence;
in the actual process, the learning process is almost the same as the standard gradient reduction, and during testing, whether the image is a clean receiving environment picture is verified firstly; if the image is predicted to be a non-clean image, the system will take a new image to continue the process until a clean image is found within a given time; second, if the image is predicted to be a vegetable category, the model will filter other labels in the vegetable in the given order; finally, through the Priori Tree CNN model combined with the information of orders, weight and the like, the clean image can obtain the food material category and the predicted score.
The formula expression mode of the primer-Tree CNN model is as follows:
firstly, supposing that a prior tree is provided, all food categories are composed of three layers of trees, as shown in fig. 2, the prior tree has K +1 leaf nodes, and corresponding to K food category labels and 1 neglected label respectively; the leaf nodes are connected with 7 father nodes of the second layer, the K food material categories have 4 fathers, namely are divided into 4 groups (S1, S2, S3 and S4), and the receiving environment categories are divided into 3 groups (N1, N2 and N3); for the classification, if the input images are classified into vegetables with high confidence level, other categories in the order are filtered for prediction, the categories such as meat, aquatic products and the like do not need to be concerned, and the accuracy rate of the food material category identification task is further improved;
in order to simplify the Priori Tree model, only the relationships between the labels in the K class and the four parent nodes thereof are concerned; before the softmax function, the leaf tag node k is associated with a weight vector
Figure BDA0001628267910000061
Associating; each super class node s and vector
Figure BDA0001628267910000062
Associating, wherein s belongs to {1, 2, 3, 4 }; e.g. betacabbageAnd betacarrotRecord and alphavegatableThe deviation therebetween; the following generative model is defined for β:
Figure BDA0001628267910000063
this prior expression describes the relationship between classes, with the conditional distribution over k being expressed as:
Figure BDA0001628267910000064
inferring the value of { W, β, α } from MAP, maximizing:
logP(k|I，t，W，β)+logP(k|I，W，β)+logP(W)+logP(β|α)+logP(α)； (3)
from the point of view of the loss function, minimize:
Figure BDA0001628267910000065
here, by fixing the value of α to 0, the loss function will be reduced to the standard loss function, let
CsThen { k | paramt (k) ═ s }, then
Figure BDA0001628267910000071
Therefore, the loss function in equation (4) is optimized by iteratively performing the following two steps; first, W and β are maximized, α is fixed using a standard random gradient descent (SGD) of a standard loss function; second, β is fixed by equation (5), thereby maximizing α.
Example 1:
training by taking a ResNet-50 multitask experiment as a baseline experiment, and then gradually improving the identification precision of food materials by using an order, a weight characteristic and a transfer learning technology, wherein a Prior-Tree CNN model has the following specific experimental method:
A. the data set is compared with the experimental environment,
the data set originates from Mealcome, which is one of the largest food supply chain platforms in china (www.mealcome.com), serving nearly 1000 restaurants; 15020 pictures of a clean environment and 15557 dirty pictures are selected to construct a Meal-53 data set; the clean pictures are divided into 51 food material categories, and the dirty pictures are divided into 3 categories; all of these data tags are added manually; in the part of the clean picture, the image numbers of each category vary from 106 to 895; for all dirty pictures, "light dark" is merged to "others", so the dirty picture is divided into two parts, the "bag not opened" and "others", the number of pictures is 11382 and 4157 respectively; the ratio of the training set, the test set and the validation set was 70%, 15% and 15%, respectively; due to the imbalance of the data set, in training the ResNet model, the method of oversample was used to ensure that there were 500 images for training and 100 images for testing per class; to train the multitask-based deep CNN model, clean and dirty picture datasets are used;
the clean pictures in the Meal-53 data set contain order information, the number of the order information is 5026, and the number of the food material pictures of each order is between 10 and 35, which is also unbalanced; in each order, the food material image also has weight information; 5026 orders are further divided into a training set, a testing set and a verification set, and the proportion of the three orders is the same as that of the data set; these data sets are used to train Order-based models (Order Dropout, Order weighing and priority Tree), with Order and weight information integrated into the image labels during training, testing and validation;
in the experiment, the depth learning frame uses the modified cafe, the depth learning model selects ResNet-50, and the ResNet-50 is pre-trained to reduce the training time and improve the accuracy; in all models, the weight lr _ mult of the last Innerproduct layer is set to 10.0, and the bias term is set to 20.0 for training; the training and testing batches are both 16, and the momentum is 0.9; all experiments were performed on Intel (R) i7CPU, 32GB RAM and GeForce GTX 1080 Ti;
B. a baseline (baseline) and an assessment method,
a deep learning model ResNet and a multitask CNN are used as baseline to complete two tasks simultaneously;
the data shown in fig. 3 is average accuracy, and for the receiving environment recognition task, the recognition ratio of clean pictures and dirty pictures can be obtained: an integration ratio and a non-integration ratio; these two ratios are True Positive Rate (TPR); for the food material category identification task, the probability of each food material category can be obtained, and the evaluation formulas of Top-k hit rate and recall rate are as follows:
Figure BDA0001628267910000081
Figure BDA0001628267910000082
in the formula, NiI ∈ (0, 1...., 50) is the number of the ith food material category, NkiIs the probability of top-k for the ith food material in the test set, n equals 50;
C. the results of the experiments and the analysis thereof,
the experimental results of ResNet, Baseline, Order weighing Model and priority Tree Model are shown in FIGS. 3 and 4, and FIG. 3 shows the experimental results obtained using the 1-crop validation method; it can be concluded that the priority Tree Model further increases the integration rate and non-integration rate by 94.41% and 92.86%, respectively, in the receiving environment recognition task; similarly, FIG. 4 shows that the Priori Tree Model achieved the best results of 99.12% and 99.94% on Top-1 and Top-3 accuracy, and that Top-5 achieved substantially consistent results of 100% on all other models; the experimental results in FIG. 4 were obtained by the 10-crop verification method, and it can be seen that the method has a better effect than the 1-crop verification method, in which it can be seen that the integer ratio and non-integer ratio in the Priori Tree Model are further improved to 94.84% and 94.56%, respectively; fig. 5 shows the recognition results of the Baseline and the Priori Tree models for some food materials, wherein the "Baseline" label under the sample image represents the recognition result of the Baseline (Baseline), and the "Priori Tree" represents the recognition result of the prior Tree Model; the above results indicate that the use of the Priori Tree CNN of the transfer learning can improve the accuracy of food material identification.
According to the large-scale food material image classification method based on the transfer learning, when only a small number of labeled samples exist in a food material data set, a convolutional neural network model (CNN) is improved based on the transfer learning method, a Prior-Tree CNN model is provided for learning food material class structures and classifier parameters, a large amount of existing labeled data of related classes are fully utilized, efficient knowledge transfer of the related classes is learned, and therefore the accuracy and robustness of object classification of a small sample set are improved by using a trained classifier on the large sample set; establishing a Meal-53 small sample data set, wherein the data set comprises 32 vegetable categories, 15 meat and poultry eggs, 3 aquatic food materials, 1 bean food material and a bean food material, and 3 dirty pictures, namely, no bag opening, dark light and the like;
the convolutional neural network model (CNN) is improved based on a transfer learning method, and a primer-Tree CNN model is provided for learning class structures and classifier parameters, and the improved model has the following advantages: the method has high efficiency, provides a multitask CNN model to classify honest/dishonest environment images so as to reduce noise in a Meal-53 data set and more effectively utilize the characteristics of training samples; the accuracy is high, and the improved model can correctly identify the category when the data training set is less; generally, the improved CNN model solves the problem through learning, is not limited to a specific problem, and can automatically build a model according to the problem, thereby solving similar problems.
The above-described embodiments are merely preferred embodiments of the present invention, and all equivalent changes or modifications of the structures, features and principles described in the claims of the present invention are included in the scope of the present invention.

Claims (1)

1. A large-scale food material image classification method based on transfer learning is characterized by comprising the following steps: inputting an original picture, and outputting information containing two aspects through the transfer learning knowledge, wherein one is the identification result of the receiving environment, and the other is the identification result of the food material type; after forward propagation of the multitask-based prior tree CNN model, the joint features produced by the last fully-connected layer will be input into different classification tasks;
the classification method specifically comprises the following steps: adding an artificially set tree structure on the multi-task classification CNN model, firstly setting a parent node according to the tree structure, and then enabling a newly added subclass to directly obtain the characteristics of the parent; after optimizing the joint loss layer and sharing the visual features, the CNN model will deliver its relevant parameters in the back propagation process; marking whether the collected food material images are pictures in a clean goods receiving environment by using c e to {0, 1}, and marking the food material types by using a multi-type label K to e to { 1.. K }, wherein K is the number of all the types; finally, training the whole network and parameters iteratively until convergence;
when testing, firstly, verifying whether the image is a clean receiving environment picture; if the image is predicted to be a non-clean image, the system will take a new image to continue the process until a clean image is found within a given time; second, if the image is predicted to be a vegetable category, the model will filter other labels in the vegetable in the given order; finally, through a priori tree CNN model combined with order and weight information, the clean image can obtain the food material category and the predicted score;
the prior tree CNN model has the following formula expression mode:
firstly, supposing that a prior tree is provided, all food categories are composed of three layers of trees, the prior tree has K +1 leaf nodes, and K food category labels and 1 neglected label correspond to the prior tree respectively; the leaf nodes are connected with 7 father nodes of the second layer, the K food material categories have 4 fathers and comprise S1, S2, S3 and S4, and the receiving environment categories are divided into 3 groups comprising N1, N2 and N3; for the hierarchical classification, if the input images are classified into vegetables with high confidence, other categories in the order are filtered out for prediction, and the categories such as meat, aquatic products and the like do not need to be concerned;
before the softmax function, the leaf tag node k is associated with a weight vector
Figure FDA0003020539080000021
Associating; each super class node s and vector
Figure FDA0003020539080000022
Associating, wherein s belongs to {1, 2, 3, 4 }; e.g. betacabbageAnd betacarrotRecord and alphavegatableThe deviation therebetween; the following generative model is defined for β:
Figure FDA0003020539080000023
this prior expression describes the relationship between classes, with the conditional distribution over k being expressed as:
Figure FDA0003020539080000024
inferring the value of { W, β, α } from MAP, maximizing:
logP(k|I，t，W，β)+logP(k|I，W，β)+logP(W)+logP(β|α)+logP(α)； (3)
from the point of view of the loss function, minimize:
Figure FDA0003020539080000025
here, by fixing the value of α to 0, the loss function will be reduced to the standard loss function, let CsThen { k | paramt (k) ═ s }, then
Figure FDA0003020539080000026
Therefore, the loss function in equation (4) is optimized by iteratively performing the following two steps; first, maximize W and β, fix α using standard random gradient descent of a standard loss function; second, β is fixed by equation (5), thereby maximizing α.
CN201810332217.0A 2018-04-13 2018-04-13 Large-scale food material image classification method based on transfer learning Active CN108647702B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810332217.0A CN108647702B (en) 2018-04-13 2018-04-13 Large-scale food material image classification method based on transfer learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810332217.0A CN108647702B (en) 2018-04-13 2018-04-13 Large-scale food material image classification method based on transfer learning

Publications (2)

Publication Number Publication Date
CN108647702A CN108647702A (en) 2018-10-12
CN108647702B true CN108647702B (en) 2021-06-01

Family

ID=63746074

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810332217.0A Active CN108647702B (en) 2018-04-13 2018-04-13 Large-scale food material image classification method based on transfer learning

Country Status (1)

Country Link
CN (1) CN108647702B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109409442A (en) * 2018-11-21 2019-03-01 电子科技大学 Convolutional neural networks model selection method in transfer learning
CN110309856A (en) * 2019-05-30 2019-10-08 华为技术有限公司 Image classification method, neural network training method and device
CN110598636B (en) * 2019-09-09 2023-01-17 哈尔滨工业大学 A Ship Target Recognition Method Based on Feature Migration
CN110807472B (en) * 2019-10-12 2022-08-12 北京达佳互联信息技术有限公司 Image recognition method and device, electronic equipment and storage medium
CN112001315B (en) * 2020-08-25 2024-01-19 中国人民解放军海军军医大学第一附属医院 Bone marrow cell classification and identification method based on migration learning and image texture characteristics
CN112257761A (en) * 2020-10-10 2021-01-22 天津大学 A Machine Learning-Based Method for Analyzing Food Nutrients in Images
CN112200249A (en) * 2020-10-13 2021-01-08 湖南大学 Autonomous updating solution of unmanned intelligent container

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105894041A (en) * 2016-04-26 2016-08-24 国网山东省电力公司经济技术研究院 Method of extracting substation information in power distribution system based on hyperspectral remote sensing images
CN107391772A (en) * 2017-09-15 2017-11-24 国网四川省电力公司眉山供电公司 A kind of file classification method based on naive Bayesian
CN107563439A (en) * 2017-08-31 2018-01-09 湖南麓川信息科技有限公司 A kind of model for identifying cleaning food materials picture and identification food materials class method for distinguishing
CN107679582A (en) * 2017-10-20 2018-02-09 深圳市唯特视科技有限公司 A kind of method that visual question and answer are carried out based on multi-modal decomposition model

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9949714B2 (en) * 2015-07-29 2018-04-24 Htc Corporation Method, electronic apparatus, and computer readable medium of constructing classifier for disease detection

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105894041A (en) * 2016-04-26 2016-08-24 国网山东省电力公司经济技术研究院 Method of extracting substation information in power distribution system based on hyperspectral remote sensing images
CN107563439A (en) * 2017-08-31 2018-01-09 湖南麓川信息科技有限公司 A kind of model for identifying cleaning food materials picture and identification food materials class method for distinguishing
CN107391772A (en) * 2017-09-15 2017-11-24 国网四川省电力公司眉山供电公司 A kind of file classification method based on naive Bayesian
CN107679582A (en) * 2017-10-20 2018-02-09 深圳市唯特视科技有限公司 A kind of method that visual question and answer are carried out based on multi-modal decomposition model

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
"Simultaneous Deep Transfer Across Domains and Tasks";Eric Tzeng et al;《 IEEE International Conference on Computer Vision》;20151213;第4068-4076页 *
"基于Tchebichef不变矩与SVM交通标志分类算法研究";丁淑艳等;《机电一体化》;20100430;摘要，引言，第3.1-3.2节，附图1-3 *
"基于数据挖掘的软件缺陷预测技术";马樱等;《厦门大学出版社》;20171231;第34-35页 *
"基于类决策树分类的特征层融合识别算法";尚朝轩等;《控制与决策》;20160630;论文第1-2节 *

Also Published As

Publication number Publication date
CN108647702A (en) 2018-10-12

Similar Documents

Publication Publication Date Title
CN108647702B (en) Large-scale food material image classification method based on transfer learning
CN104992142B (en) A kind of pedestrian recognition method being combined based on deep learning and attribute study
Hussein et al. Videograph: Recognizing minutes-long human activities in videos
Zhong et al. Re-ranking person re-identification with k-reciprocal encoding
Khan et al. Real-time plant health assessment via implementing cloud-based scalable transfer learning on AWS DeepLens
Meyers et al. Im2Calories: towards an automated mobile vision food diary
CN107301380A (en) One kind is used for pedestrian in video monitoring scene and knows method for distinguishing again
Zhang et al. Unsupervised difference representation learning for detecting multiple types of changes in multitemporal remote sensing images
CN109583322A (en) A kind of recognition of face depth network training method and system
Chen et al. Modeling fashion
Cong et al. Self-supervised online metric learning with low rank constraint for scene categorization
CN109492075B (en) Transfer learning sequencing method based on loop generation countermeasure network
CN108009560B (en) Commodity image similarity category judgment method and device
Mac Aodha et al. Putting the scientist in the loop--Accelerating scientific progress with interactive machine learning
Zhang et al. Learning object detectors with semi-annotated weak labels
Jenrette et al. Shark detection and classification with machine learning
Shoohi et al. DCGAN for Handling Imbalanced Malaria Dataset based on Over-Sampling Technique and using CNN.
CN110992217A (en) Method and device for expressing and searching multi-view features of design patent
CN108664986A (en) Based on lpThe multi-task learning image classification method and system of norm regularization
AU2019100969A4 (en) Chinese Food Recognition and Search System
CN112183464A (en) Video pedestrian identification method based on deep neural network and graph convolution network
Chang et al. Fine-grained butterfly and moth classification using deep convolutional neural networks
Bai et al. Learning two-pathway convolutional neural networks for categorizing scene images
CN110569761A (en) A Hand-drawn Sketch Retrieval Method for Remote Sensing Images Based on Adversarial Learning
CN110263808A (en) A kind of Image emotional semantic classification method based on LSTM network and attention mechanism

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