CN113537325A - Deep learning method for image classification based on logic of extracting high-low-level features - Google Patents

Deep learning method for image classification based on logic of extracting high-low-level features Download PDF

Info

Publication number
CN113537325A
CN113537325A CN202110758190.3A CN202110758190A CN113537325A CN 113537325 A CN113537325 A CN 113537325A CN 202110758190 A CN202110758190 A CN 202110758190A CN 113537325 A CN113537325 A CN 113537325A
Authority
CN
China
Prior art keywords
logic
feature
extraction unit
layer
feature extraction
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.)
Granted
Application number
CN202110758190.3A
Other languages
Chinese (zh)
Other versions
CN113537325B (en
Inventor
马辛
付幸文
孙亦琦
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.)
Beihang University
Original Assignee
Beihang 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 Beihang University filed Critical Beihang University
Priority to CN202110758190.3A priority Critical patent/CN113537325B/en
Publication of CN113537325A publication Critical patent/CN113537325A/en
Application granted granted Critical
Publication of CN113537325B publication Critical patent/CN113537325B/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
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a deep learning method based on extracting high-low level feature logic for image classification, which consists of a feature extraction network and a logic network, wherein the feature extraction network is obtained by connecting a plurality of feature extraction units in series, the logic network is obtained by connecting a plurality of logic extraction units in series, the feature extraction units are convolutional layers or residual blocks, the logic extraction units are convolutional dendritic modules consisting of convolutional layers and Hadamard products, the feature extraction network is single-input and multi-output, is used for extracting image features of all levels from low to high, inputting images to be classified and outputting feature maps of all levels, the logic network is multi-input and single-output, and is used for constructing logic relations among high-low level image features and classifying according to the logic relations, inputting feature maps of all levels and outputting classification results Faster convergence speed and better robustness.

Description

Deep learning method for image classification based on logic of extracting high-low-level features
Technical Field
The invention relates to the field of artificial intelligence, in particular to a deep learning method for extracting high-low-level feature logic for classification, which is applied to image signals.
Background
Classification is a fundamental problem in many areas of fault diagnosis, automation, Computer Vision (CV), and Natural Language Processing (NLP). Deep learning has been a useful tool to solve the classification problem. Feature extraction has long been the key to Convolutional Neural Networks (CNNs) to mimic human vision. The CNN performs feature extraction on the input image by a convolution kernel. Due to the ability of translation and other degeneration and hierarchical expression, the CNN can extract deeper image features for classification. Later, researchers improved the CNN at depth with skip-junctions to get a residual neural network ResNet. The model solves the problems of gradient disappearance, gradient diffusion and degradation of deep CNN, so that the neural network has stronger feature extraction capability. The biological dendrites in the brain are proved to have logical operation ability of AND \ OR \ NOT, and the logical relation between input signals can be calculated. The existing model for simulating the function of the biological dendrites is a tree-projection network (DD) which can provide measures for logical extraction, and as the number of layers is increased, the DD can construct more complex logical relations. The convolutional dendritic network (CDD) inherits the logic extraction capability of the DD and the feature extraction capability of the CNN, and overcomes the defects that the DD cannot perform feature extraction and the CNN cannot perform logic extraction. Thus, CDD focuses more on the combination of features and relies on the logical relationship between features for classification. Many models in the field of target detection, such as a Feature Pyramid Network (FPN), DensNet, YOLOv3, and the like, improve the detection accuracy of small objects by fusing high-level and low-level features. In the field of image segmentation, a large number of semantic structures are kept in a high-level semantic segmentation result, but a small structure is seriously lost, a large number of details are kept in a low-level feature segmentation result, but semantic category prediction is poor, so that a semantic segmentation model with better performance is obtained by fusing high-level and low-level features and combining the advantages of the high-level and low-level features.
There have been many attempts to mimic the human visual perception capabilities, but the following problems still remain. The traditional neural network model cannot perform logic extraction between features. Although the CDD has the advantages of DD and CNN, the CDD has a shallow feature extraction layer and thus has a weak feature extraction capability. In addition, because of the hierarchical expression capability of the neural network on the features, both the high-level features and the low-level features contain a large amount of information of the image, and the existing classification network only utilizes the high-level features for classification and loses a large amount of information of the low-level features.
Disclosure of Invention
The invention aims to solve the technical problems that the traditional network cannot perform logic extraction among features, the feature extraction capability of a convolutional dendritic network (CDD) is weak, and the existing neural network cannot combine high-and-low-layer features, and provides a deep learning method for image classification based on logic for extracting the high-and-low-layer features. The method is mainly applied to the field of image classification, can improve the classification precision of all traditional models for classification only by extracting image features, and has the advantages of high classification precision, high convergence speed, good robustness and the like.
The technical scheme adopted by the invention for solving the technical problems is as follows: a deep learning method based on logic of extracting high-level and low-level features for image classification is characterized by comprising a feature extraction network and a logic network, wherein the feature extraction network is obtained by connecting a plurality of feature extraction units in series, the logic network is obtained by connecting a plurality of logic extraction units in series, the feature extraction units are convolutional layers or residual blocks, the logic extraction units are convolutional dendritic modules consisting of convolutional layers and Hadamard products, the connection relationship among the units is that the input of the feature extraction unit at the layer is the output of the feature extraction unit at the upper level, the input of the logic extraction unit at the layer is the output of the logic extraction unit at the upper level and the output of the feature extraction unit at the layer, the feature extraction network is single-input and multi-output and is used for extracting the image features of all levels from low level to high level, the input is the image to be classified and the feature map of all levels is output, the logic network is multi-input single-output and is used for constructing the logic relation among the high-bottom image characteristics and classifying according to the logic relation, the input is a characteristic diagram of each layer, and the output is a classification result, and the method specifically comprises the following steps:
step 1, sending an input image with an image label into a first-layer convolutional layer for dimension conversion and primary feature extraction to obtain a feature map 0;
Figure BDA0003148029510000021
in the formula (I), the compound is shown in the specification,
Figure BDA0003148029510000026
is a characteristic diagram 0, XinFor inputting an image, WfirstIs the weight matrix of the first convolutional layer,
Figure BDA0003148029510000022
is a convolution;
step 2, sending the feature map 0 into a feature extraction unit 1 to obtain a feature map 1 after feature extraction;
Figure BDA0003148029510000023
in the formula (I), the compound is shown in the specification,
Figure BDA0003148029510000027
is a characteristic diagram 0, XF 1Is characterized in FIG. 1, WF 1Weight matrix, W, for feature extraction unit 1S 1For the linear mapping weight matrix in the hop connection in the feature extraction unit 1, F (-) is a non-linear activation function,
Figure BDA0003148029510000024
is a convolution;
step 3, simultaneously sending the feature map 0 and the feature map 1 to a logic extraction unit 1, and obtaining a feature logic map 1 after high-low layer feature logic combination;
Figure BDA0003148029510000025
in the formula (I), the compound is shown in the specification,
Figure BDA0003148029510000031
in order to characterize the logic diagram of figure 1,
Figure BDA00031480295100000325
is a characteristic diagram 0, XF 1In order to characterize the drawing of figure 1,
Figure BDA0003148029510000032
is the weight matrix of the logical extraction unit 1,
Figure BDA0003148029510000033
for convolution, O is the Hadamard product;
step 4, sending the feature map 1 into a feature extraction unit 2, and obtaining a feature map 2 after feature extraction;
Figure BDA0003148029510000034
in the formula, XF 1In order to characterize the drawing of figure 1,
Figure BDA00031480295100000320
in order to characterize the figure 2 in the way,
Figure BDA00031480295100000321
a weight matrix, W, for the feature extraction unit 2S 2For the linear mapping weight matrix in the hop connection in the feature extraction unit 2, F (-) is a non-linear activation function,
Figure BDA0003148029510000035
is a convolution;
step 5, simultaneously sending the characteristic logic diagram 1 and the characteristic diagram 2 into a logic extraction unit 2, and obtaining a characteristic logic diagram 2 after high-low layer characteristic logic combination;
Figure BDA0003148029510000036
in the formula (I), the compound is shown in the specification,
Figure BDA0003148029510000037
in order to characterize the logic diagram of figure 1,
Figure BDA0003148029510000038
for the characteristic logic diagram 2, XF 1Is a characteristic diagram of FIG. 1, XF 2In order to characterize the figure 2 in the way,
Figure BDA0003148029510000039
is the weight matrix of the logic extraction unit 2,
Figure BDA00031480295100000310
for convolution, O is the Hadamard product;
and 6, continuously utilizing the next-layer feature extraction unit to extract the features of a higher layer and utilizing the next-layer logic extraction unit to logically combine the features of a higher layer and a lower layer until a feature logic diagram of a highest layer is obtained, wherein a recurrence formula is expressed as follows:
Figure BDA00031480295100000311
Figure BDA00031480295100000312
wherein k is a layer mark (k: 2,3, …),
Figure BDA00031480295100000322
in order to characterize the diagram k-1,
Figure BDA00031480295100000323
in order to be able to characterize the graph k,
Figure BDA00031480295100000313
in order to characterize the logic diagram k-1,
Figure BDA00031480295100000314
in order to characterize the logic diagram k,
Figure BDA00031480295100000324
a weight matrix, W, for the feature extraction unit kS kFor the linear mapping weight matrix in the hopping connection in the feature extraction unit k,
Figure BDA00031480295100000315
for the weight matrix of the logic extraction unit 2, F (-) is a non-linear activation function,
Figure BDA00031480295100000316
is a convolution;
step 7, reforming the feature logic diagram of the highest layer into a one-dimensional tensor and sending the one-dimensional tensor into a classifier, and obtaining a classification result of the image by the classifier according to the feature logic of the highest layer;
Figure BDA00031480295100000317
in the formula, XoutIn order to be a result of the classification,
Figure BDA00031480295100000318
is a characteristic logic diagram of the highest layer, WFCReshape (-) represents the reformation of variables into one-dimensional tensors for the weight matrix of the fully connected layers in the classifier;
step 8, comparing the classification result with the image label, calculating a classification error, and adjusting the weights of the first-layer convolutional layer, each feature extraction unit and each logic extraction unit by using an error back propagation algorithm to iterate for multiple times to obtain an optimal image classification model;
E＝L(Xlab,Xout)
Figure BDA00031480295100000319
in the formula, XoutFor the classification result, XlabIs an image label, E is a classification error, W*For all weights in the model, W*' is all weights of the model adjusted by the error back propagation algorithm, L (-) is a loss function, and eta is a learning rate;
and 9, classifying the input images without the image labels by using the optimal image classification model to obtain the categories of the input images.
Further, expanding the formula in step 3 to obtain a feature logic diagram 1 including all possible feature logic combinations between the feature diagram 0 and the feature diagram 1;
Figure BDA0003148029510000041
Figure BDA0003148029510000042
Figure BDA0003148029510000043
in the formula (I), the compound is shown in the specification,
Figure BDA0003148029510000044
in order to characterize the logic diagram of figure 1,
Figure BDA0003148029510000048
is a characteristic diagram 0, XF 1In order to characterize the drawing of figure 1,
Figure BDA0003148029510000045
is the weight of logical extraction unit 1, fij 0Is the image feature of the feature map 0, fij 1For characterizing the image features of FIG. 1, and logic fij 0·fij 1OR logic fij 0+fij 1NOT logic-fij *，
Figure BDA0003148029510000046
Is the weight of the logical extraction unit 1, where i, j is the matrix index (i ═ 1,2, …, n;. j ═ 1,2, …, m), n is the length of the image, m is the width of the image, x represents any layer,
Figure BDA0003148029510000047
for convolution, o is the hadamard product, (padding 1, stop 1) indicates that one pixel is filled around the image and the convolution step size is 1.
Further, by using the recursion formula in step 6, the feature logic combination among all the high-level and low-level features can be obtained, and then the weight before the logic combination is adjusted by an error back propagation algorithm, so that the high-level and low-level feature logic relationship contributing to the classification precision can be obtained.
The principle of the invention is as follows: traditional classification networks can extract more robust image features for classification, but ignore potential logical relationships between features. Meanwhile, the classification mode only depends on the high-level features, and a large amount of detail features contained in the low-level features are ignored. Therefore, the invention combines the feature extraction module with better network robustness and the CDD module through a well-designed framework, extracts and combines the features of the high-bottom image, and finally classifies the features according to the logical relationship among the features.
Compared with the prior art, the invention has the advantages that:
the advantages compared to conventional models (such as ResNet, WRN, etc.) that are used for classification only by extracting image features are: the invention considers the logic relation between the high-bottom characteristics, and the classification precision and the convergence speed are both obviously improved; the advantages compared to CDD networks that rely on logical relationships between features for classification are: the method has stronger feature extraction capability, and can extract image features of higher levels, so that the model is more robust.
Drawings
FIG. 1 is a flow chart of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, exemplary embodiments of the present invention are described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
In order to illustrate the deep learning method for extracting high-level and low-level feature logic for classification, a complete image classification model is constructed by the method, image signals of two source data sets Cifar10 and Cifar100 are classified, and the classification target related in the technical scheme can be tensors with any dimension and size, and is compared with two traditional models (ResNet, WRN) which only rely on high-level features for classification under the same frame. And the model applied to ResNet is called ResNet-CDD, and the model applied to WRN is called WRN-CDD.
As shown in fig. 1, the present invention is composed of a feature extraction unit and a logic extraction unit. And (3) replacing the input data with 3 multiplied by 32 tensors according to the sizes of the images in the Cifar10 and Cifar100 data, quantizing the images, sending the quantized images into the model, training the model according to the steps 1 to 8, and obtaining the optimal classification model after multiple iterations. The whole process of extracting the characteristics and obtaining the beneficial characteristic logic combination aims at reducing errors, and automatic iterative optimization is carried out by taking an error back propagation algorithm as a method without manual participation. The model prediction classification process is steps 1 to 7. Since the weights of various types in the model are adjusted and optimized by a large amount of data in the training set, the error between the output of the image signals in the test set obtained through the model and the image labels is also within an allowable range.
As shown in tables 1,2, 3 and 4, in order to compare the present invention with two conventional models (ResNet, WRN) classified only by means of high-level features on two open source data sets Cifar10 and Cifar100 in the same framework, the model applied to ResNet is called ResNet-CDD, and the model applied to WRN is called WRN-CDD.
The following phenomena can be seen from tables 1,2, 3 and 4: on the Cifar10 data set with small classification difficulty, when the original model is small (the number of layers or width or the number of parameters is small, namely the feature extraction capability is weak), the precision of the original model is improved greatly, and when the model is gradually enlarged (namely the feature extraction capability is gradually enhanced), the precision of the original model gradually reaches the precision level of the method. On the Cifar100 data set with high classification difficulty, the method greatly improves the original model, and also presents a trend that the improvement of the original model is changed from large to small as the original model is changed from small to large (namely the feature extraction capability of the original model is changed from small to large).
TABLE 1 comparative experimental results of the data set for Cifar10 between the present invention (ResNet-CDD) and the original model (ResNet)
Figure BDA0003148029510000061
TABLE 2 comparative experimental results of the Cifar100 dataset for the present invention (ResNet-CDD) and the former model (ResNet)
Figure BDA0003148029510000062
TABLE 3 comparative experimental results of the data set of Cifar10 between the present invention (WRN-CDD) and the original model (WRN)
Figure BDA0003148029510000071
TABLE 4 comparative experimental results of the present invention (WRN-CDD) and the original model (WRN) in the Cifar100 dataset
Figure BDA0003148029510000081
First, it can be seen from the above that the present invention has the advantages of higher precision, faster convergence speed, and better robustness. Secondly, the invention can greatly enhance the classification model with weak feature extraction capability, which accords with the visual characteristics of human beings, in fact, human beings can realize high-precision classification only by extracting some low-level image features (such as contour, texture, color and the like) and then through the logic combination of the brain, and human beings can not be limited to only carry out logic combination on the features of the same level but can combine all available high-level features. Meanwhile, if the recognized object is simple and only needs a few simple features to distinguish, human vision does not need to classify through logic combination among the features, which explains that the original model can achieve the same level of classification accuracy on the Cifar10 data set as the present invention.
Those skilled in the art will appreciate that the invention may be practiced without these specific details.
The above examples are provided only for the purpose of describing the present invention, and are not intended to limit the scope of the present invention. The scope of the invention is defined by the appended claims. Various equivalent substitutions and modifications can be made without departing from the spirit and principles of the invention, and are intended to be within the scope of the invention.

Claims (1)

1. A deep learning method based on logic of extracting high-level and low-level features for image classification is characterized by comprising a feature extraction network and a logic network, wherein the feature extraction network is obtained by connecting a plurality of feature extraction units in series, the logic network is obtained by connecting a plurality of logic extraction units in series, the feature extraction units are convolutional layers or residual blocks, the logic extraction units are convolutional dendritic modules consisting of convolutional layers and Hadamard products, the connection relationship among the units is that the input of the feature extraction unit at the layer is the output of the feature extraction unit at the upper level, the input of the logic extraction unit at the layer is the output of the logic extraction unit at the upper level and the output of the feature extraction unit at the layer, the feature extraction network is single-input and multi-output and is used for extracting the image features of all levels from low level to high level, the input is the image to be classified and the feature map of all levels is output, the logic network is multi-input single-output and is used for constructing the logic relation among the high-bottom image characteristics and classifying according to the logic relation, the input is a characteristic diagram of each layer, and the output is a classification result, and the method specifically comprises the following steps:
step 1, sending an input image with an image label into a first-layer convolutional layer for dimension conversion and primary feature extraction to obtain a feature map 0;
Figure FDA0003148029500000011
in the formula, XF 0Is a characteristic diagram 0, XinFor inputting an image, WfirstIs the weight matrix of the first convolutional layer,
Figure FDA0003148029500000012
is a convolution;
step 2, sending the feature map 0 into a feature extraction unit 1 to obtain a feature map 1 after feature extraction;
Figure FDA0003148029500000013
in the formula, XF 1Is characterized in FIG. 1, WF 1Weight matrix, W, for feature extraction unit 1S 1For the linear mapping weight matrix in the hop connection in the feature extraction unit 1, F (-) is a non-linear activation function,
Figure FDA0003148029500000014
is a convolution;
step 3, simultaneously sending the feature diagram 0 and the feature diagram 1 to a logic extraction unit 1, and obtaining a feature logic diagram 1 after high-level and low-level feature logic combination, wherein the feature logic diagram 1 comprises all possible feature logic combinations between the feature diagram 0 and the feature diagram 1;
Figure FDA0003148029500000021
Figure FDA0003148029500000022
Figure FDA0003148029500000023
in the formula (I), the compound is shown in the specification,
Figure FDA0003148029500000024
for the characteristic logic diagram 1, XF 0Is a characteristic diagram 0, XF 1In order to characterize the drawing of figure 1,
Figure FDA0003148029500000025
is the weight of logical extraction unit 1, fij 0Is the image feature of the feature map 0, fij 1For characterizing the image features of FIG. 1, and logic fij 0·fij 1OR logic fij 0+fij 1NOT logic-fij *，
Figure FDA0003148029500000026
Is the weight of the logical extraction unit 1, where i, j is the matrix index (i ═ 1,2, …, n;. j ═ 1,2, …, m), n is the length of the image, m is the width of the image, x represents any layer,
Figure FDA0003148029500000027
in order to be a convolution,
Figure FDA0003148029500000028
is the product of Hadamard;
step 4, sending the feature map 1 into a feature extraction unit 2, and obtaining a feature map 2 after feature extraction;
Figure FDA0003148029500000029
in the formula, XF 2Is characterized by FIG. 2, WF 2A weight matrix, W, for the feature extraction unit 2S 2Linear mapping weight matrix in jump connection in the feature extraction unit 2;
step 5, simultaneously sending the characteristic logic diagram 1 and the characteristic diagram 2 into a logic extraction unit 2, and obtaining a characteristic logic diagram 2 after high-low layer characteristic logic combination;
Figure FDA00031480295000000210
in the formula (I), the compound is shown in the specification,
Figure FDA00031480295000000211
in order to characterize the logic of figure 2,
Figure FDA00031480295000000212
is the weight matrix of the logic extraction unit 2;
and 6, continuously extracting the features of a higher layer by using a next-layer feature extraction unit and logically combining the features of a high layer and a low layer by using a next-layer logic extraction unit until a feature logic diagram of the highest layer is obtained, wherein the feature logic diagram comprises feature logic combinations among all the features of the high layer and the low layer, and a recursion formula is expressed as follows:
Figure FDA00031480295000000213
Figure FDA00031480295000000214
wherein k is a layer number, k is 2,3, …, XF k-1Is a characteristic diagram k-1, XF kIn order to be able to characterize the graph k,
Figure FDA0003148029500000031
in order to characterize the logic diagram k-1,
Figure FDA0003148029500000032
is a characteristic logic diagram k, WF kA weight matrix, W, for the feature extraction unit kS kLinear mapping weight matrix in jump connection in the feature extraction unit k;
step 7, reforming the feature logic diagram of the highest layer into a one-dimensional tensor and sending the one-dimensional tensor into a classifier, and obtaining a classification result of the image by the classifier according to the feature logic of the highest layer;
Figure FDA0003148029500000033
in the formula, XoutIn order to be a result of the classification,
Figure FDA0003148029500000034
is a characteristic logic diagram of the highest layer, WFCReshape (-) represents the reformation of variables into one-dimensional tensors for the weight matrix of the fully connected layers in the classifier;
step 8, comparing the classification result with the image label, calculating a classification error, adjusting the weights of the first-layer convolutional layer, each feature extraction unit and each logic extraction unit by using an error back propagation algorithm, and obtaining an optimal image classification model after multiple iterations;
E＝L(Xlab,Xout)
Figure FDA0003148029500000035
in the formula, XoutFor the classification result, XlabIs an image label, E is a classification error, W*For all weights in the model, W*' is all weights of the model adjusted by the error back propagation algorithm, L (-) is a loss function, and eta is a learning rate;
and 9, classifying the input images without the image labels by using the optimal image classification model to obtain the categories of the input images.
CN202110758190.3A 2021-07-05 2021-07-05 Deep learning method for image classification based on extracted high-low layer feature logic Active CN113537325B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110758190.3A CN113537325B (en) 2021-07-05 2021-07-05 Deep learning method for image classification based on extracted high-low layer feature logic

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110758190.3A CN113537325B (en) 2021-07-05 2021-07-05 Deep learning method for image classification based on extracted high-low layer feature logic

Publications (2)

Publication Number Publication Date
CN113537325A true CN113537325A (en) 2021-10-22
CN113537325B CN113537325B (en) 2023-07-11

Family

ID=78126781

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110758190.3A Active CN113537325B (en) 2021-07-05 2021-07-05 Deep learning method for image classification based on extracted high-low layer feature logic

Country Status (1)

Country Link
CN (1) CN113537325B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114886440A (en) * 2022-07-13 2022-08-12 武汉工程大学 Epileptic sample discharge classification model training and recognition method, system and electronic equipment

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020216227A1 (en) * 2019-04-24 2020-10-29 华为技术有限公司 Image classification method and apparatus, and data processing method and apparatus
WO2021057810A1 (en) * 2019-09-29 2021-04-01 深圳数字生命研究院 Data processing method, data training method, data identifying method and device, and storage medium
CN112926691A (en) * 2021-04-07 2021-06-08 北京航空航天大学 Convolutional dendrite method for extracting feature logic for classification

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020216227A1 (en) * 2019-04-24 2020-10-29 华为技术有限公司 Image classification method and apparatus, and data processing method and apparatus
WO2021057810A1 (en) * 2019-09-29 2021-04-01 深圳数字生命研究院 Data processing method, data training method, data identifying method and device, and storage medium
CN112926691A (en) * 2021-04-07 2021-06-08 北京航空航天大学 Convolutional dendrite method for extracting feature logic for classification

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
陈瑞瑞;: "基于深度卷积神经网络的图像分类算法", 河南科技学院学报(自然科学版), no. 04 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114886440A (en) * 2022-07-13 2022-08-12 武汉工程大学 Epileptic sample discharge classification model training and recognition method, system and electronic equipment

Also Published As

Publication number Publication date
CN113537325B (en) 2023-07-11

Similar Documents

Publication Publication Date Title
CN111368896B (en) Hyperspectral remote sensing image classification method based on dense residual three-dimensional convolutional neural network
CN109711426B (en) Pathological image classification device and method based on GAN and transfer learning
CN108399421B (en) Deep zero sample classification method based on word embedding
CN108399428B (en) Triple loss function design method based on trace ratio criterion
CN111242288B (en) Multi-scale parallel deep neural network model construction method for lesion image segmentation
CN109783666B (en) Image scene graph generation method based on iterative refinement
CN110516539A (en) Remote sensing image building extracting method, system, storage medium and equipment based on confrontation network
CN112784801A (en) Text and picture-based bimodal gastric disease classification method and device
CN109389171B (en) Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology
CN113240683B (en) Attention mechanism-based lightweight semantic segmentation model construction method
CN112950780B (en) Intelligent network map generation method and system based on remote sensing image
CN108595558B (en) Image annotation method based on data equalization strategy and multi-feature fusion
CN108615231A (en) A kind of full reference picture assessment method for encoding quality based on neural network learning fusion
CN104036242A (en) Object recognition method based on convolutional restricted Boltzmann machine combining Centering Trick
CN113255602A (en) Dynamic gesture recognition method based on multi-modal data
CN113537325A (en) Deep learning method for image classification based on logic of extracting high-low-level features
CN114170657A (en) Facial emotion recognition method integrating attention mechanism and high-order feature representation
Lonij et al. Open-world visual recognition using knowledge graphs
CN114295967A (en) Analog circuit fault diagnosis method based on migration neural network
CN111639751A (en) Non-zero padding training method for binary convolutional neural network
CN116993639A (en) Visible light and infrared image fusion method based on structural re-parameterization
Di et al. FDNet: An end-to-end fusion decomposition network for infrared and visible images
CN114494284B (en) Scene analysis model and method based on explicit supervision area relation
CN113486929B (en) Rock slice image identification method based on residual shrinkage module and attention mechanism
CN115063374A (en) Model training method, face image quality scoring method, electronic device and storage medium

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