CN111340096A - Weakly supervised butterfly target detection method based on confrontation complementary learning - Google Patents

Weakly supervised butterfly target detection method based on confrontation complementary learning Download PDF

Info

Publication number
CN111340096A
CN111340096A CN202010111404.3A CN202010111404A CN111340096A CN 111340096 A CN111340096 A CN 111340096A CN 202010111404 A CN202010111404 A CN 202010111404A CN 111340096 A CN111340096 A CN 111340096A
Authority
CN
China
Prior art keywords
branch
butterfly
network
image
classifier
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202010111404.3A
Other languages
Chinese (zh)
Inventor
李玉鑑
方宇
张婷
刘兆英
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.)
Beijing University of Technology
Original Assignee
Beijing University of Technology
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 Beijing University of Technology filed Critical Beijing University of Technology
Priority to CN202010111404.3A priority Critical patent/CN111340096A/en
Publication of CN111340096A publication Critical patent/CN111340096A/en
Pending legal-status Critical Current

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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/53—Querying
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • 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
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features

Landscapes

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

Abstract

The invention discloses a weak supervision butterfly target detection method based on antagonistic complementary learning, which sequentially comprises the following steps: firstly, mixing a butterfly ecological image crawled by a crawler with a butterfly specimen image according to categories to form a butterfly data set; then cutting and standardizing the image; then dividing the butterfly data set into a training image set and a testing image set according to a proportion; then, a backbone network and an confrontation complementary learning network are established, a training set is used for training the network, and the model is stored when the network converges; and finally, inputting the test image into the trained network model to obtain a target detection result graph.

Description

Weakly supervised butterfly target detection method based on confrontation complementary learning
Technical Field
The invention belongs to the technical field of image recognition, and particularly relates to a weak supervision butterfly target detection method based on confrontation complementary learning.
Background
Butterflies are a group of insects of the phylum arthropoda, class insecta, order lepidoptera, suborder hammer. On one hand, butterflies in the larval stage feed on agricultural and forestry crops and are one of main pests in the agricultural and forestry; on the other hand, the butterfly is a precious environmental index, the monitoring data of the butterfly is used for ecological environment monitoring, biodiversity protection and the like, and meanwhile, the butterfly has high ornamental value and economic value and is a natural resource. Therefore, the classification and identification of butterflies have great significance in the aspects of agriculture, forestry, disease and pest control, environmental protection, butterfly industry development and other practical works.
The traditional butterfly identification mode mainly comprises two modes of artificial identification and biochemical identification. The manual identification mode mainly compares the ecological characteristics with the specimen characteristics, and the method depends on long-term experience accumulation and is time-consuming; the biochemical identification method utilizes the reaction of butterfly genitals to biochemical reagents, and the method depends on professional biochemical knowledge and is expensive. Therefore, the two methods do not have the universality of butterfly identification.
With the development of image processing technology and machine learning theory, researchers realize identification of butterflies through a machine learning method, which mainly comprises the steps of artificially extracting image features (color, texture and shape information of butterfly wing surfaces) of butterflies, performing mathematical modeling according to the feature information, and determining a classifier for classification.
Most of machine learning methods need to manually select image features, and the final effect of classification is determined to a great extent by feature extraction and feature selection. Meanwhile, machine learning methods focus on identification of butterfly specimen images, and an effective identification means for butterfly ecological images (butterfly images shot in an ecological environment) is lacking. In the ecological image, on one hand, most butterflies do not occupy all positions in the image; on the other hand, butterflies have mimicry capability, so that butterfly targets are difficult to distinguish from backgrounds, which brings huge challenges to identification of butterfly images. Therefore, in order to better identify the butterfly in the ecological image, the position of the butterfly in the image needs to be determined, and then the identification of the butterfly is completed, which is the target detection of the butterfly.
Deep learning methods, represented by Convolutional Neural Networks (CNNs), have been highly successful in the field of image recognition. The deep learning method can automatically extract image features, and makes great breakthrough in various tasks such as image classification, target detection, image segmentation and the like. Aiming at a target detection task, a two-stage target detection algorithm represented by an R-CNN series and a one-stage target detection algorithm represented by SSD, YOLO and the like have excellent effects, but the algorithms belong to full-supervision detection algorithms, depend on manually marked object boundary boxes and are expensive. To solve the problem of fully supervised detection algorithms, researchers have come to focus on how to accomplish target detection under weak supervision (image-level label only), and some success has emerged. For example, Zhou et al uses a Global Average Pooling (GAP) layer instead of the fully connected layer of VGG to obtain the position information of the object, but this method can only obtain the most discriminative region. Singh et al randomly conceals the block of the region from each input image based on the Zhou method to obtain more distinct regions. However, this method cannot effectively locate the entire area of the object due to random concealment.
The invention effectively locates all positions of the butterfly in the image in a weak supervision mode by using the image-level label only through the confrontation complementary learning method and identifies the category of the butterfly.
Disclosure of Invention
The invention aims to provide a weak supervision butterfly target detection method based on confrontation complementary learning, which is used for target detection of butterfly images. In order to achieve the purpose, the invention adopts the following technical scheme:
a weak supervision butterfly target detection method based on antagonistic complementary learning comprises the following steps:
step 1: and constructing a butterfly data set. The butterfly data set is composed of two parts, the first part is composed of Google pictures and butterfly ecological images crawled on Baidu pictures, and the first part is called as a data set D1The second part is composed of the butterfly specimen image on "Chinese butterfly Zhi", which is calledData set D2. Data set D1And a data set D2Mixed composition of butterfly data sets
Figure BDA0002390134660000021
Wherein the butterfly image is IiThe category label is yi. The data set D contains N images of M butterflies in total, and the data set D is divided into a training set Dt(containing N)tImage) and test set Ds(containing N)sA web image);
step 2: and constructing a backbone network. The present invention selects the first 13 layers of the VGG-16 as the backbone network, which consists of 5 convolutional blocks. Butterfly image I with color input for backbone networki∈Rh×w×3(1<i<Nt) Where h and w represent the height and width of the image, respectively, and 3 represents the number of channels of the image. Location-aware feature maps with multi-channel output for networks
Figure BDA0002390134660000022
Wherein K1Number of channels, H, representing a location-aware feature map1And W1Respectively, the height and width of the feature map. The backbone network is represented as:
Si＝f0(θ0,Ii)
wherein f is0(. -) represents the role of the backbone network, θ0Is a parameter of the backbone network;
and step 3: and constructing an antagonistic complementary learning network. The countervailing complementary learning network comprises two parallel branches A and B, each branch comprising a feature extractor and a classifier. Wherein the feature extractor and classifier of the A branch are respectively denoted as EAAnd clsAThe feature extractor and classifier of the B branch are denoted EBAnd clsB；
Step 3.1: for the A branch, the branch first uses a feature extractor EAExtracting features and acquiring a category activation graph; then uses the classifier clsAAnd (6) classifying. Wherein the feature extractor is a three-layer convolutional neural network whose input is boneOutput S of the trunk networki(1≤i≤Nt) Output as class target map
Figure BDA0002390134660000031
The target graph shows the unique distinguished regions of the target class. Will be provided with
Figure BDA0002390134660000032
Normalized to [0,1 ]]And is defined as
Figure BDA0002390134660000033
Figure BDA0002390134660000034
The location map of the branch is obtained. The classifier consists of a Global Average Pooling (GAP) layer and a soft max (softmax) layer, whose inputs are
Figure BDA0002390134660000035
Output as classification result
Figure BDA0002390134660000036
Where M represents the number of categories of butterflies. The whole branch is specifically represented as:
Figure BDA0002390134660000037
Figure BDA0002390134660000038
wherein f isA(. a) and
Figure BDA0002390134660000039
respectively represent feature extractor EAAnd a classifier clsAAction of thetaAAs a feature extractor EAIs determined by the parameters of (a) and (b),
Figure BDA00023901346600000310
parameters of the A branch classifier;
step 3.2: erasing feature maps using a feature eraser Era
Figure BDA00023901346600000311
The most discriminative region in the set. Assuming that the threshold is δ, the most discriminating region is
Figure BDA00023901346600000312
The area where Aera is located is in the characteristic diagram SiSetting the middle value to be 0, and generating a feature diagram after erasing
Figure BDA00023901346600000313
Namely, it is
Figure BDA00023901346600000314
Step 3.3: for the B branch, the structure is substantially the same as for the A branch. The branch first uses a feature extractor EBExtracting features, obtaining class activation graph, and then using classifier clsBAnd (6) classifying. Wherein, the feature extractor is also a three-layer convolution neural network, and the input of the feature extractor is an erased feature map
Figure BDA0002390134660000041
Output as class target map
Figure BDA0002390134660000042
Figure BDA0002390134660000043
The new most discriminative area will be learned. Will be provided with
Figure BDA0002390134660000044
Normalized to [0,1 ]]And is defined as
Figure BDA0002390134660000045
Figure BDA0002390134660000046
The location map of the branch is obtained. The classifier consists of a global average pooling layer anda soft maximum layer having as its input
Figure BDA0002390134660000047
Output as classification result
Figure BDA0002390134660000048
The whole branch is specifically represented as:
Figure BDA0002390134660000049
Figure BDA00023901346600000410
wherein f isB(. a) and
Figure BDA00023901346600000411
respectively represent feature extractor EBAnd a classifier clsBAction of thetaBAs a feature extractor EBIs determined by the parameters of (a) and (b),
Figure BDA00023901346600000412
parameters of the B-branch classifier;
and 4, step 4: establishing a loss function L of two branch networksAAnd LBThe loss function being the actual output vector
Figure BDA00023901346600000413
And
Figure BDA00023901346600000414
respectively with the target output vector yiRespectively expressed as:
Figure BDA00023901346600000415
Figure BDA00023901346600000416
then netThe total loss of the collaterals is L ═ LA+LB；
And 5: and (5) network training. Setting super parameters such as iteration times, learning rate and the like, and setting a training set DtInputting a network, iteratively updating network parameters by using a random gradient descent algorithm until loss is converged, and storing a final model;
step 6: and (5) testing the network. Loading the saved model, and testing the set DsAnd inputting the data into a network to obtain the accuracy of classification. Inputting a single test image Ii∈Rh×w×3(1≤i≤Ns) Obtaining the location map of the A branch
Figure BDA00023901346600000417
And location map of branch B
Figure BDA00023901346600000418
Taking the maximum value of the corresponding positions of the two positioning maps to obtain the final positioning map
Figure BDA0002390134660000051
And drawing a rectangular frame on the image according to the positioning diagram, so as to obtain the position of the butterfly target in the image.
Drawings
Fig. 1 is an original image.
Fig. 2 is a backbone network structure.
Fig. 3 shows the overall network structure.
Fig. 4 is a graph of the test results.
Detailed Description
The embodiment of the invention provides a weak supervision butterfly target detection method based on antagonistic complementary learning, and the invention is explained and explained below by combining the related drawings:
the flow of the embodiment of the invention is as follows:
step 1: and constructing a butterfly data set. The butterfly data set consists of two parts, namely a butterfly ecological image data set and a butterfly specimen image data set. The butterfly ecological image is obtained by crawling reptiles on Google pictures and Baidu pictures and is called as a data set D1The number ofThe image of the data set is shown in fig. 1 (a). The butterfly specimen image is obtained from the book "Chinese butterfly log", and is called as data set D2The image of the data set is shown in fig. 1 (b). Data set D1And a data set D2Mixed composition of butterfly data sets
Figure BDA0002390134660000052
Wherein the butterfly image is IiThe category label is yiThe butterfly data set D is classified into M334, which contains N74111 images. Dividing the data set D into training sets D according to the proportion of 8:2 of each classt(containing N)t58288 images) and test set Ds(containing N)s14823 images) to prevent the computational burden, each butterfly image is resampled to 256 × 256 and then randomly cropped to 224 × 224 as input to the network, where the data needs to be normalized (the dimensions of the image minus the mean of the dataset and divided by the standard deviation of the dataset);
step 2: and constructing a backbone network. The present invention selects the first 13 layers of the VGG-16 as the backbone network, which consists of 5 convolutional blocks. Wherein, for the first 2 convolutional blocks, each convolutional block consists of 2 convolutional layers; for the last 3 convolutional blocks, each convolutional block is composed of 3 convolutional layers, and the structure is shown in fig. 2, wherein the total number of the convolutional layers is 13. Butterfly image I with color input for backbone networki∈R224×224×3(1<i<Nt) Where 3 denotes the number of image channels, and h-224 and w-224 denote the height and width of the image, respectively. Location-aware feature map S with multi-channel output for networki∈R28×28×512Where 512 represents the number of channels of the feature map, 28 × 28 represents the resolution of the feature map, the backbone network is represented as:
Si＝f0(θ0,Ii),1＜i＜Nt
wherein f is0(. -) represents the role of the backbone network, θ0Is a parameter of the backbone network;
and step 3: and constructing an antagonistic complementary learning network. The antagonistic complementary learning network comprises two of A and BThe line branches, each branch containing a feature extractor and a classifier. Wherein the feature extractor and classifier of the A branch are respectively denoted as EAAnd clsAThe feature extractor and classifier of the B branch are denoted EBAnd clsB；
Step 3.1: for the first branch A, the branch first uses a feature extractor EAExtracting features, obtaining class activation map, and using classifier clsAAnd (6) classifying. Wherein the feature extractor is a three-layer convolutional neural network, and the input of the convolutional neural network is the output S of the backbone networkiOutput as class target map
Figure BDA0002390134660000061
The figure shows the unique distinguishing region of the target class. Will be provided with
Figure BDA0002390134660000062
Normalized to [0,1 ]]And is defined as
Figure BDA0002390134660000063
Figure BDA0002390134660000064
The location map of the branch is obtained. The classifier consists of a Global Average Pooling (GAP) layer and a soft max layer. The global average pooling layer replaces a full connection layer in the VGG16, and the output of the global average pooling layer is a one-dimensional vector with the size of 334; the soft max layer maps one-dimensional vectors to probabilities for each class. The input of the classifier is
Figure BDA0002390134660000065
Output as classification result
Figure BDA0002390134660000066
The whole branch is specifically represented as:
Figure BDA0002390134660000067
wherein f isA(. a) and
Figure BDA0002390134660000068
respectively represent feature extractor EAAnd a classifier clsAAction of thetaAAs a feature extractor EAIs determined by the parameters of (a) and (b),
Figure BDA0002390134660000069
parameters of the A branch classifier;
step 3.2: erasing feature maps using a feature eraser Era
Figure BDA00023901346600000610
Assuming the threshold is δ (δ ∈ {0.5,0.6,0.7,0.8,0.9}), the most discriminative region is
Figure BDA00023901346600000611
The area where Aera is located is in the characteristic diagram SiSetting the middle value to be 0, and generating a feature diagram after erasing
Figure BDA0002390134660000071
Namely, it is
Figure BDA0002390134660000072
Step 3.3: for the second branch B, the structure is substantially the same as branch a. The branch first uses a feature extractor EBExtracting features, obtaining class activation graph, and then using classifier clsBAnd (6) classifying. Wherein, the feature extractor is also a three-layer convolution neural network, and the input of the feature extractor is an erased feature map
Figure BDA0002390134660000073
Output as class target map
Figure BDA0002390134660000074
Figure BDA0002390134660000075
Will learn the most new judgmentA sexual area. Will be provided with
Figure BDA0002390134660000076
Normalized to [0,1 ]]And is defined as
Figure BDA0002390134660000077
Figure BDA0002390134660000078
The location map of the branch is obtained. The classifier consists of a global average pooling layer and a soft maximum layer, wherein the output of the global average pooling layer is a one-dimensional vector with the size of 334; the soft max layer maps one-dimensional vectors to probabilities for each class. The input of the classifier is
Figure BDA0002390134660000079
Output as classification result
Figure BDA00023901346600000710
The whole branch is specifically represented as:
Figure BDA00023901346600000711
wherein f isB(. a) and
Figure BDA00023901346600000712
respectively represent feature extractor EBAnd a classifier clsBAction of thetaBAs a feature extractor EBIs determined by the parameters of (a) and (b),
Figure BDA00023901346600000713
parameters of the B-branch classifier;
and 4, step 4: establishing a loss function L of two branch networksAAnd LBThe loss function being the actual output vector
Figure BDA00023901346600000714
And
Figure BDA00023901346600000715
and the target output vector yiRespectively expressed as:
Figure BDA00023901346600000716
Figure BDA00023901346600000717
the total loss of the network is L ═ LA+LB；
And 5: and (5) network training. Setting the iteration number to be 50, the learning rate to be 0.001 and the threshold value delta to be 0.6, and setting the training set DtInputting a network, initializing a backbone network by using a VGG16 weight trained by ImageNet, iteratively updating network parameters by using a random gradient descent algorithm until loss is converged, and storing a final model;
step 6: and (5) testing the network. Loading the saved model, and testing the set DsAnd inputting the data into a network to obtain the classification accuracy. Inputting a single test image Ii∈Rh×w×3(1≤i≤Ns) Obtaining the location map of the A branch
Figure BDA0002390134660000081
And location map of branch B
Figure BDA0002390134660000082
Taking the maximum value of the corresponding positions of the two positioning maps to obtain the final positioning map
Figure BDA0002390134660000083
Adding the scout map to the test image, as shown in fig. 4 (a); and (3) binarizing the positioning map, then obtaining the contour of the butterfly target, then obtaining a circumscribed rectangle of the contour, and finally drawing the rectangle on the test image to obtain the position of the butterfly target in the image, as shown in (b) of the attached figure 4.
The above examples are only used to describe the present invention, and do not limit the technical solutions described in the present invention. Therefore, all technical solutions and modifications that do not depart from the spirit and scope of the present invention should be construed as being included in the scope of the appended claims.

Claims (2)

1. A weak supervision butterfly target detection method based on confrontation complementary learning is characterized by comprising the following steps: the method comprises the following steps of,
step 1: constructing a butterfly data set: the butterfly data set is composed of two parts, the first part is composed of Google pictures and butterfly ecological images crawled on Baidu pictures, and the first part is called as a data set D1The second part consisting of an image of a butterfly specimen, called dataset D2(ii) a Data set D1And a data set D2Mixed composition of butterfly data sets
Figure FDA0002390134650000011
Wherein the butterfly image is IiThe category label is yi(ii) a The data set D contains N images of M butterflies in total, and the data set D is divided into a training set DtAnd test set Ds(ii) a Training set DtContaining NtImage, test set DsContaining NsA frame of images;
step 2: constructing a backbone network: selecting the first 13 layers of the VGG-16 as a backbone network, wherein the backbone network consists of 5 convolution blocks; butterfly image I with color input for backbone networki∈Rh×w×3，1<i<NtWherein h and w respectively represent the height and width of the image, and 3 represents the number of channels of the image; location-aware feature maps with multi-channel output for networks
Figure FDA0002390134650000012
Wherein K1Number of channels, H, representing a location-aware feature map1And W1The height and width of the feature map, respectively; the backbone network is represented as:
Si＝f0(θ0,Ii),
wherein f is0(. -) represents the role of the backbone network, θ0Is a parameter of the backbone network;
and step 3: construction of antagonistic complementationLearning the network: the countervailing complementary learning network comprises two parallel branches A and B, wherein each branch comprises a feature extractor and a classifier; wherein the feature extractor and classifier of the A branch are respectively denoted as EAAnd clsAThe feature extractor and classifier of the B branch are denoted EBAnd clsB；
And 4, step 4: establishing a loss function L of two branch networksAAnd LBThe loss function being the actual output vector
Figure FDA0002390134650000013
And
Figure FDA0002390134650000014
respectively with the target output vector yiRespectively expressed as:
Figure FDA0002390134650000015
Figure FDA0002390134650000016
the total loss of the network is L ═ LA+LB；
And 5: network training: setting super parameters such as iteration times, learning rate and the like, and setting a training set DtInputting a network, iteratively updating network parameters by using a random gradient descent algorithm until loss is converged, and storing a final model;
step 6: network testing: loading the saved model, and testing the set DsInputting the data into a network to obtain the accuracy of classification; inputting a single test image Ii∈Rh×w×3Obtaining the location map of the A branch
Figure FDA0002390134650000021
And location map of branch B
Figure FDA0002390134650000022
Taking the maximum value of the corresponding positions of the two positioning maps to obtain the final positioning map
Figure FDA0002390134650000023
And drawing a rectangular frame on the image according to the positioning diagram to obtain the position of the butterfly target in the image.
2. The method for detecting a weakly supervised butterfly target based on antagonistic complementary learning as claimed in claim 1, wherein: step 3 comprises the following steps, step 3.1: for the A branch, the branch first uses a feature extractor EAExtracting features and acquiring a category activation graph; then uses the classifier clsAClassifying; wherein the feature extractor is a three-layer convolutional neural network, and the input of the convolutional neural network is the output S of the backbone networkiOutputting the class target map
Figure FDA0002390134650000024
The target graph shows a unique distinguished region of the target class; will be provided with
Figure FDA0002390134650000025
Normalized to [0,1 ]]And is defined as
Figure FDA0002390134650000026
Figure FDA0002390134650000027
The positioning diagram of the branch is obtained; classifier clsAComprises a global average pooling layer and a soft maximum output layer, and the input is
Figure FDA0002390134650000028
Output as classification result
Figure FDA0002390134650000029
The whole branch is specifically represented as:
Figure FDA00023901346500000210
Figure FDA00023901346500000211
wherein f isA(. a) and
Figure FDA00023901346500000212
respectively represent feature extractor EAAnd a classifier clsAAction of thetaAAs a feature extractor EAIs determined by the parameters of (a) and (b),
Figure FDA00023901346500000213
parameters of the A branch classifier;
step 3.2: erasing feature maps using a feature eraser Era
Figure FDA00023901346500000214
The most discriminative region: assuming the threshold is δ, the most discriminative region can be represented as
Figure FDA00023901346500000215
The Area where the Area is located is positioned in a characteristic diagram SiSetting the middle value to be 0, and generating a feature diagram after erasing
Figure FDA00023901346500000216
Namely, it is
Figure FDA00023901346500000217
Step 3.3: for the B branch, the structure is basically the same as that of the A branch; the branch also first uses a feature extractor EBExtracting features, obtaining class activation map, and using classifier clsBClassifying; wherein, the feature extractor is also a three-layer convolution neural network, and the input is the erased feature map
Figure FDA0002390134650000031
Output as class target map
Figure FDA0002390134650000032
Figure FDA0002390134650000033
Learning a new most discriminative area; will be provided with
Figure FDA0002390134650000034
Normalized to [0,1 ]]And is defined as
Figure FDA0002390134650000035
Figure FDA0002390134650000036
The positioning diagram of the branch is obtained; the classifier consists of a global average pooling layer and a soft max layer with inputs of
Figure FDA0002390134650000037
Output as classification result
Figure FDA0002390134650000038
The whole branch is specifically represented as:
Figure FDA0002390134650000039
Figure FDA00023901346500000310
wherein f isB(. a) and
Figure FDA00023901346500000311
respectively represent feature extractor EBAnd a classifier clsBAction of thetaBAs a feature extractor EBIs determined by the parameters of (a) and (b),
Figure FDA00023901346500000312
are parameters of the B-branch classifier.
CN202010111404.3A 2020-02-24 2020-02-24 Weakly supervised butterfly target detection method based on confrontation complementary learning Pending CN111340096A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010111404.3A CN111340096A (en) 2020-02-24 2020-02-24 Weakly supervised butterfly target detection method based on confrontation complementary learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010111404.3A CN111340096A (en) 2020-02-24 2020-02-24 Weakly supervised butterfly target detection method based on confrontation complementary learning

Publications (1)

Publication Number Publication Date
CN111340096A true CN111340096A (en) 2020-06-26

Family

ID=71185486

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010111404.3A Pending CN111340096A (en) 2020-02-24 2020-02-24 Weakly supervised butterfly target detection method based on confrontation complementary learning

Country Status (1)

Country Link
CN (1) CN111340096A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112364980A (en) * 2020-11-09 2021-02-12 北京计算机技术及应用研究所 Deep neural network training method based on reinforcement learning under weak supervision scene
CN112801029A (en) * 2021-02-09 2021-05-14 北京工业大学 Multi-task learning method based on attention mechanism
CN112800927A (en) * 2021-01-25 2021-05-14 北京工业大学 AM-Softmax loss-based butterfly image fine granularity identification method
CN114882298A (en) * 2022-07-11 2022-08-09 东声(苏州)智能科技有限公司 Optimization method and device for confrontation complementary learning model

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105488536A (en) * 2015-12-10 2016-04-13 中国科学院合肥物质科学研究院 Agricultural pest image recognition method based on multi-feature deep learning technology
CN109063742A (en) * 2018-07-06 2018-12-21 平安科技（深圳）有限公司 Butterfly identifies network establishing method, device, computer equipment and storage medium
CN109376765A (en) * 2018-09-14 2019-02-22 汕头大学 A kind of butterfly automatic classification method based on deep learning
CN109886295A (en) * 2019-01-11 2019-06-14 平安科技（深圳）有限公司 A kind of butterfly recognition methods neural network based and relevant device
CN110569901A (en) * 2019-09-05 2019-12-13 北京工业大学 Channel selection-based countermeasure elimination weak supervision target detection method

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105488536A (en) * 2015-12-10 2016-04-13 中国科学院合肥物质科学研究院 Agricultural pest image recognition method based on multi-feature deep learning technology
CN109063742A (en) * 2018-07-06 2018-12-21 平安科技（深圳）有限公司 Butterfly identifies network establishing method, device, computer equipment and storage medium
CN109376765A (en) * 2018-09-14 2019-02-22 汕头大学 A kind of butterfly automatic classification method based on deep learning
CN109886295A (en) * 2019-01-11 2019-06-14 平安科技（深圳）有限公司 A kind of butterfly recognition methods neural network based and relevant device
CN110569901A (en) * 2019-09-05 2019-12-13 北京工业大学 Channel selection-based countermeasure elimination weak supervision target detection method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
XIAOLIN ZHANG 等: "Adversarial Complementary Learning for Weakly Supervised Object Localization", 《2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 *

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112364980A (en) * 2020-11-09 2021-02-12 北京计算机技术及应用研究所 Deep neural network training method based on reinforcement learning under weak supervision scene
CN112364980B (en) * 2020-11-09 2024-04-30 北京计算机技术及应用研究所 Deep neural network training method based on reinforcement learning under weak supervision scene
CN112800927A (en) * 2021-01-25 2021-05-14 北京工业大学 AM-Softmax loss-based butterfly image fine granularity identification method
CN112800927B (en) * 2021-01-25 2024-03-29 北京工业大学 Butterfly image fine-granularity identification method based on AM-Softmax loss
CN112801029A (en) * 2021-02-09 2021-05-14 北京工业大学 Multi-task learning method based on attention mechanism
CN112801029B (en) * 2021-02-09 2024-05-28 北京工业大学 Attention mechanism-based multitask learning method
CN114882298A (en) * 2022-07-11 2022-08-09 东声(苏州)智能科技有限公司 Optimization method and device for confrontation complementary learning model
CN114882298B (en) * 2022-07-11 2022-11-01 东声(苏州)智能科技有限公司 Optimization method and device for confrontation complementary learning model

Similar Documents

Publication Publication Date Title
CN109214399B (en) Improved YOLOV3 target identification method embedded in SENET structure
Xie et al. Multilevel cloud detection in remote sensing images based on deep learning
CN109977918B (en) Target detection positioning optimization method based on unsupervised domain adaptation
CN111340096A (en) Weakly supervised butterfly target detection method based on confrontation complementary learning
Grilli et al. A review of point clouds segmentation and classification algorithms
Li et al. SAR image change detection using PCANet guided by saliency detection
CN110796186A (en) Dry and wet garbage identification and classification method based on improved YOLOv3 network
CN110619059B (en) Building marking method based on transfer learning
CN109766873B (en) Pedestrian re-identification method based on hybrid deformable convolution
CN112464911A (en) Improved YOLOv 3-tiny-based traffic sign detection and identification method
CN111833322B (en) Garbage multi-target detection method based on improved YOLOv3
CN113269224B (en) Scene image classification method, system and storage medium
CN112101364B (en) Semantic segmentation method based on parameter importance increment learning
CN113761259A (en) Image processing method and device and computer equipment
CN112528845B (en) Physical circuit diagram identification method based on deep learning and application thereof
CN111339935A (en) Optical remote sensing picture classification method based on interpretable CNN image classification model
CN111753682A (en) Hoisting area dynamic monitoring method based on target detection algorithm
CN111353396A (en) Concrete crack segmentation method based on SCSEOCUnet
CN111079837A (en) Method for detecting, identifying and classifying two-dimensional gray level images
CN112132145A (en) Image classification method and system based on model extended convolutional neural network
CN111414951B (en) Fine classification method and device for images
CN113920472A (en) Unsupervised target re-identification method and system based on attention mechanism
Tan et al. Rapid fine-grained classification of butterflies based on FCM-KM and mask R-CNN fusion
CN115049952A (en) Juvenile fish limb identification method based on multi-scale cascade perception deep learning network
CN112308825A (en) SqueezeNet-based crop leaf disease identification method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20200626

WD01 Invention patent application deemed withdrawn after publication