CN111860672B - Fine-grained image classification method based on block convolutional neural network - Google Patents

Fine-grained image classification method based on block convolutional neural network Download PDF

Info

Publication number
CN111860672B
CN111860672B CN202010738474.1A CN202010738474A CN111860672B CN 111860672 B CN111860672 B CN 111860672B CN 202010738474 A CN202010738474 A CN 202010738474A CN 111860672 B CN111860672 B CN 111860672B
Authority
CN
China
Prior art keywords
block
convolution
fine
feature map
neural network
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
CN202010738474.1A
Other languages
Chinese (zh)
Other versions
CN111860672A (en
Inventor
马占宇
谢吉洋
杜若一
司中威
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 Posts and Telecommunications
Original Assignee
Beijing University of Posts and Telecommunications
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 Posts and Telecommunications filed Critical Beijing University of Posts and Telecommunications
Priority to CN202010738474.1A priority Critical patent/CN111860672B/en
Publication of CN111860672A publication Critical patent/CN111860672A/en
Application granted granted Critical
Publication of CN111860672B publication Critical patent/CN111860672B/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
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • 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

Abstract

A fine-grained image classification method based on a block convolutional neural network relates to the technical field of fine-grained image identification, solves the problem that the existing method inputs an original image into the convolutional neural network after being averagely blocked for fine-grained image classification, and has weak reception field limitation. The invention limits the convolution receptive field according to the requirement, so that the network focuses more on the characteristics of the local area and is more suitable for the fine-grained image classification task. The fine-grained image classification method limits the receptive field range of the convolutional layer on the premise of not introducing more parameters, so that a convolutional neural network can search a smaller discriminative local area.

Description

Fine-grained image classification method based on block convolutional neural network
Technical Field
The invention relates to the technical field of fine-grained image recognition, in particular to a fine-grained image classification method based on a block convolutional neural network.
Background
In the technical Field of fine-grained image recognition, most of the existing methods based on artificial intelligence and deep learning directly input images into a Convolutional Neural Network (CNN), a Feature Map is extracted from an output Feature Map (Feature Map) of a previous layer through multilayer convolution and pooling layer operation, a Feature Map with a larger Receptive Field (RF), namely a range in which each Feature point on the Feature Map is mapped onto an input image, is extracted layer by layer, and finally a Feature Map with the Receptive Field being the size of the whole image (the theoretical Receptive Field may be larger than the size of the whole image) is obtained and used for fine-grained image classification. However, most existing methods are mainly used to identify the type of object in the image, such as different colored wings and different shaped beaks in birds, different shaped lights and tires in automobiles, by finding discriminative local areas on the image. In this case, the smaller field of view enables the model to better extract local features on the image, and thus to search for smaller discriminative local regions. However, the existing convolutional neural network framework mainly introduces operations with higher complexity and larger parameter amount, but still has difficulty in limiting the receptive field size of the convolutional layer.
Fine-Grained Visual Classification (FGVC) is a sub-task of the traditional image Classification task, which refers to more refined Classification of objects of a certain class, for example: distinguish different kinds of birds or dogs, different models of automobiles or airplanes, and the like. The fine-grained degree is more challenging than the traditional classification task because the difference between the target object and the objects of the same category may be larger than the difference between the target object and the objects of different categories, for example, two birds of the same category may have a great difference due to different postures; however, two birds of different species may have differences in structure and texture only in local areas such as the beak and the tail of a bird due to their close physical statures.
With the development of deep learning, CNN has become a mainstream solution for the task of image classification. CNN is mainly composed of the following parts: (1) a convolutional layer for feature extraction; (2) the pooling layer is used for feature selection and information filtering; (3) and the full connection layer is used for carrying out nonlinear combination on the extracted features to obtain the final output. In CNN, the concept of RF refers to a range in which a feature point on an output feature map of a specified layer is mapped onto an input picture, and both a convolutional layer and a pooling layer have the effect of increasing the receptive field, and the receptive field relationship between adjacent layers of a network is calculated in the following manner:
Figure BDA0002605832620000021
wherein r is(l)The reception field, k, of the first layer of the convolutional or pooling layer(l)Refers to the kernel size, s, of the first convolutional or pooling layer(l′)Refers to the step size of the i' th convolutional or pooling layer.
The existing fine-grained classification methods are mainly divided into two types: (1) based on the local positioning method, the convolutional neural network is required to be used for extracting features, a plurality of discriminative areas are found, the areas are cut from an original image, and feature extraction and classification operations are respectively executed, so that the prediction time is long; in addition, the number of regions for classification is mostly set in advance in the method, and the flexibility of the model is greatly limited. (2) Most of methods based on end-to-end feature coding generate a high-dimensional vector before a full connection layer to improve the model expression capability to adapt to a fine-grained classification task. The extra computation amount brought by the excessively high dimensionality greatly limits the model efficiency.
For a traditional convolutional neural network, the general receptive field is very large, and for a general image classification task, the model can be judged according to information in a larger range; however, for fine-grained tasks, too large a reception field increases the influence of intra-class differences on the network, making it difficult to focus on local details.
The existing document, "fine-grained visual classification based on jigsaw and progressive multi-grained learning" is a method that an original image is averagely blocked, disturbed and blocked and then directly input into a convolutional neural network for fine-grained image classification, and is different in that (1) the method only blocks the original image; (2) the method limits the receptive field by a method of disturbing the blocks, and the limitation is weaker.
Disclosure of Invention
The invention provides a fine-grained image classification method based on a block convolutional neural network, aiming at solving the problem that the existing method has weak receptive field limitation when an original image is input into the convolutional neural network after being averagely blocked for fine-grained image classification.
A fine-grained image classification method based on a block convolutional neural network is characterized in that the block convolutional neural network is set to have L block convolutional layers, wherein L is the number of layers of the current block convolutional layers, L is more than or equal to 1 and is less than or equal to L, and the initialization is that L is equal to 1; the method is realized by the following steps:
step one, for the first block convolution layer f (·; omega)(l)) Obtaining its input characteristic diagram as x(l)(ii) a The above-mentioned
Figure BDA0002605832620000022
For the convolution kernel parameters, R represents a real number, c(l)For inputting the number of channels of the feature map, c(l+1)For the number of channels of the output profile,. the input of the representation function,
Figure BDA0002605832620000031
and
Figure BDA0002605832620000032
for each convolution kernel width and height;
Figure BDA0002605832620000033
the dimension of expression is
Figure BDA0002605832620000034
A real matrix of (d);
the input feature map
Figure BDA0002605832620000035
Is the output characteristic diagram, x, of the (l-1) th block convolutional layer(1)As model input, W(l)And H(l)Width and height of the input feature map;
step two, when l is equal to 1, setting m1＝n 11 is ═ 1; when l > 1, the number of blocks m per line on the input feature map is calculated by the following formulalAnd the number of blocks per column nl：
Figure BDA0002605832620000036
Figure BDA0002605832620000037
In the formula (I), the compound is shown in the specification,
Figure BDA0002605832620000038
and
Figure BDA0002605832620000039
are respectively an input feature map x(l)Has a width and a height of a theoretical receptive field, and
Figure BDA00026058326200000310
and
Figure BDA00026058326200000311
is the contraction factor of the theoretical receptive field in the width and height dimensions,
Figure BDA00026058326200000312
and
Figure BDA00026058326200000313
step sizes of convolution kernels of the l' th layer of block convolution layer in the width and height dimensions of the feature map respectively,
Figure BDA00026058326200000314
the operation of rounding up is carried out;
step three, according to the number m of blocks in each row and each column on the input feature map obtained in the step twolAnd nlRandomly sampling to obtain the width of the feature map block
Figure BDA00026058326200000315
And height
Figure BDA00026058326200000316
And is
Figure BDA00026058326200000317
i＝1，…，mlAnd
Figure BDA00026058326200000318
j＝1，…，nlare all positive integers, and are not limited to the integer,
Figure BDA00026058326200000319
step four, dividing the block width according to the characteristic diagram obtained in the step three
Figure BDA00026058326200000320
And height
Figure BDA00026058326200000321
Feature map x to be input(l)Is divided into ml×nlBlock, obtaining a set of block feature maps
Figure BDA00026058326200000322
Figure BDA00026058326200000323
Step five, adopting the convolution kernel parameter omega in the step one(l)Respectively comparing all obtained in step four
Figure BDA00026058326200000324
Performing convolution to obtain corresponding convolution output characteristic diagram
Figure BDA00026058326200000325
Step six, the convolution output characteristic diagram obtained in the step five is used
Figure BDA00026058326200000326
Splicing according to the original position to obtain the output characteristic diagram of the first convolution layer in the block convolution neural network
Figure BDA00026058326200000327
Step seven, for the L block convolution layers, the operation is carried out according to the steps from one step to six until the output characteristic diagram x of the last L block convolution layer is obtained(L+1)X is to be(L+1)Inputting the data into a full connection layer to obtain the output probability p ∈ R of fine-grained image classificationnAnd n is the number of categories, so that the classification of fine-grained images is realized.
The invention has the beneficial effects that: the fine-grained image classification method can limit the experience field of convolution according to requirements, enables the network to pay more attention to the characteristics of local areas, and is more suitable for being applied to fine-grained image classification tasks. Meanwhile, additional parameters and operation are not introduced, and the high efficiency of the general convolutional neural network can be reserved in the prediction process.
The fine-grained image classification method does not need the characteristic of overlarge receptive field to divide the input characteristic graph into blocks, and each block is spliced again after being respectively subjected to convolution operation, so that the method has strong limitation.
The fine-grained image classification method limits the receptive field range of the convolutional layer on the premise of not introducing more parameters, so that a convolutional neural network can search a smaller discriminative local area.
Drawings
Fig. 1 is a flowchart of a fine-grained image classification method based on a block convolutional neural network according to the present invention.
FIG. 2 is a schematic diagram of a fine-grained image classification method based on a block convolutional neural network according to the present invention, which is expressed in ml＝nlTake 4 as an example.
FIG. 3 is a diagram of a second embodiment of a fine-grained image classification method based on a block convolutional neural network according to the present invention, where m is the numberl＝nlTake 4 as an example.
Detailed Description
In a first specific embodiment, the first embodiment is described with reference to fig. 1 and fig. 2, a fine-grained image classification method based on a block convolutional neural network is provided, where the block convolutional neural network has L block convolutional layers, L is the number of layers of the current block convolutional layer, L is greater than or equal to 1 and less than or equal to L, and is initialized to L is 1; the method is realized by the following steps:
step one, for the first block convolution layer f (·; omega)(l)) ObtainingThe input characteristic diagram is x(l)(ii) a "·" denotes an input of a function, and may be denoted by "·" when the input is uncertain. The above-mentioned
Figure BDA0002605832620000041
For the convolution kernel parameters, R is a real number,
Figure BDA0002605832620000042
the dimension of expression is
Figure BDA0002605832620000043
Real matrix of (d) for representing Ω(l)The size of (d); c. C(l)For inputting the number of channels of the feature map, c(l+1)In order to output the number of channels of the feature map,
Figure BDA0002605832620000044
and
Figure BDA0002605832620000045
for each convolution kernel width and height;
the input feature map
Figure BDA0002605832620000046
Is the output characteristic diagram, x, of the (l-1) th block convolutional layer(1)As model input, W(l)And H(l)Width and height of the input feature map;
step two, when l is equal to 1, setting m1＝n 11 is ═ 1; when l > 1, the number of blocks m per line on the input feature map is calculated by the following formulalAnd the number of blocks per column nl：
Figure BDA0002605832620000051
Figure BDA0002605832620000052
In the formula (I), the compound is shown in the specification,
Figure BDA0002605832620000053
and
Figure BDA0002605832620000054
are respectively an input feature map x(l)Has a width and a height of a theoretical receptive field, and
Figure BDA0002605832620000055
and
Figure BDA0002605832620000056
is the contraction factor of the theoretical receptive field in the width and height dimensions,
Figure BDA0002605832620000057
and
Figure BDA0002605832620000058
step sizes of convolution kernels of the l' th layer of block convolution layer in the width and height dimensions of the feature map respectively,
Figure BDA0002605832620000059
the operation of rounding up is carried out; the range of the shrinkage factor on the width dimension and the height dimension of the theoretical receptive field is respectively as follows:
Figure BDA00026058326200000510
step three, according to the number m of blocks in each row and each column on the input feature map obtained in the step twolAnd nlRandomly sampling to obtain the width of the feature map block
Figure BDA00026058326200000511
And height
Figure BDA00026058326200000512
And is
Figure BDA00026058326200000513
i＝1，…，mlAnd
Figure BDA00026058326200000514
j＝1，…，nlare all positive integers, and are not limited to the integer,
Figure BDA00026058326200000515
step four, dividing the block width according to the characteristic diagram obtained in the step three
Figure BDA00026058326200000516
And height
Figure BDA00026058326200000517
Feature map x to be input(l)Is divided into ml×nlBlock, obtaining a set of block feature maps
Figure BDA00026058326200000518
Figure BDA00026058326200000519
Step five, adopting the convolution kernel parameter omega in the step one(l)Respectively comparing all obtained in step four
Figure BDA00026058326200000520
Performing convolution to obtain corresponding convolution output characteristic diagram
Figure BDA00026058326200000521
Step six, the convolution output characteristic diagram obtained in the step five is used
Figure BDA00026058326200000522
i＝1，…，ml，j＝1，…，nlSplicing according to the original position to obtain the output characteristic diagram of the first convolution layer in the block convolution neural network
Figure BDA00026058326200000523
Step seven, for the L partitioned convolution layers, operating according to the steps one to six until the L partitioned convolution layers are all operatedObtaining the output characteristic diagram x of the last L-th block convolution layer(L+1)X is to be(L+1)Inputting the data into a full connection layer to obtain the output probability p ∈ R of fine-grained image classificationnAnd n is the number of categories, so that the classification of fine-grained images is realized.
Step eight, in the model training process, cross entropy L is usedCE(t, p) and the real category t optimize the output probability p of the fine-grained image classification:
LCE(t，p)＝-ln pt。
in a second embodiment, the present embodiment is described with reference to fig. 3, and the present embodiment is an example of a fine-grained image classification method based on a block convolutional neural network according to the first embodiment: the embodiment can simplify the operation and improve the block convolution efficiency while finishing the block convolution operation. Setting a block convolutional neural network to have L block convolutional layers, wherein L is the number of layers of the current block convolutional layer, L is more than or equal to 1 and is less than or equal to L, and initializing to L is 1;
step 1, for the first block convolution layer f (·; omega) in the block convolution neural network(l)) And "·" represents the input to the function.
Figure BDA0002605832620000061
Is the parameter of its convolution kernel, R is a real number,
Figure BDA0002605832620000062
the dimension of expression is
Figure BDA0002605832620000063
Real matrix of (d) for representing Ω(l)The size of (d); c. C(l)Is the number of channels of the input characteristic diagram,
Figure BDA0002605832620000064
and
Figure BDA0002605832620000065
is the width and height of each convolution kernel, and obtains its input characteristic diagram
Figure BDA0002605832620000066
Is the output characteristic diagram, W, of the first-1 blocked convolutional layer(l)And H(l)Is the width and height of the input feature map;
step 2, according to the number m of blocks in each row and each column on the preset feature diagramlAnd nlRandomly sampling to obtain the width of the feature map block
Figure BDA0002605832620000067
And height
Figure BDA0002605832620000068
And is
Figure BDA0002605832620000069
i＝1，…，mlAnd
Figure BDA00026058326200000610
j＝1，…，nlare all positive integers, and are not limited to the integer,
Figure BDA00026058326200000611
step 3, inputting the characteristic diagram x(l)Every other row above
Figure BDA00026058326200000612
Insert into
Figure BDA00026058326200000613
Column all zero column vector, every other row
Figure BDA00026058326200000614
Insert into
Figure BDA00026058326200000615
The rows are all zero row vectors and the row vectors,
Figure BDA00026058326200000616
and
Figure BDA00026058326200000617
is the step size of the convolution kernel in the width and height dimensions of the feature map,
Figure BDA00026058326200000618
obtaining a processed feature map for a round-down operation
Figure BDA00026058326200000619
Step 4, adopting a convolution kernel parameter omega(l)To pair
Figure BDA0002605832620000071
Performing convolution to obtain a convolution output characteristic diagram
Figure BDA0002605832620000072
Step 5, according to the positions of the all-zero column vectors and the all-zero row vectors inserted in the step 3, outputting the feature graph by convolution
Figure BDA0002605832620000073
The inserted vector is removed, and the removed column is marked with
Figure BDA0002605832620000074
Figure BDA0002605832620000075
The removed row numbers are
Figure BDA0002605832620000076
Obtaining an output feature map of the first convolutional layer in a partitioned convolutional neural network
Figure BDA0002605832620000077
Step 6, for all the block convolution layers, operating according to the steps 1 to 5 until obtaining the output characteristic diagram x of the last block convolution layer (L layer)(L+1)X is to be(L+1)Inputting the fine-grained image classification data into a full connection layer to obtain an output probability p of fine-grained image classification;
in this embodiment modeUsing Cross Entropy (CE) LCE(t, p) and the true class t optimize the output probability p of the fine-grained image classification.

Claims (4)

1. A fine-grained image classification method based on a block convolutional neural network is characterized in that the block convolutional neural network is set to have L block convolutional layers, wherein L is the number of layers of the current block convolutional layers, L is more than or equal to 1 and is less than or equal to L, and the initialization is that L is equal to 1; the method is characterized in that:
the method is realized by the following steps:
step one, for the first block convolution layer f (·; omega)(l)) Obtaining its input characteristic diagram as x(l)(ii) a The above-mentioned
Figure FDA0002605832610000011
For the convolution kernel parameters, R represents a real number, c(l)For inputting the number of channels of the feature map, c(l+1)For the number of channels of the output profile,. the input of the representation function,
Figure FDA0002605832610000012
and
Figure FDA0002605832610000013
for each convolution kernel width and height;
Figure FDA0002605832610000014
the dimension of expression is
Figure FDA0002605832610000015
A real matrix of (d);
the input feature map
Figure FDA0002605832610000016
Is the output characteristic diagram, x, of the (l-1) th block convolutional layer(1)As model input, W(l)And H(l)Width and height of the input feature map;
step two, when l is equal to 1, setting m1＝n11 is ═ 1; when in useWhen l is more than 1, calculating the number m of blocks per line on the input feature map by the following formulalAnd the number of blocks per column nl：
Figure FDA0002605832610000017
Figure FDA0002605832610000018
In the formula (I), the compound is shown in the specification,
Figure FDA0002605832610000019
and
Figure FDA00026058326100000110
are respectively an input feature map x(l)Has a width and a height of a theoretical receptive field, and
Figure FDA00026058326100000111
Figure FDA00026058326100000112
and
Figure FDA00026058326100000113
is the contraction factor of the theoretical receptive field in the width and height dimensions,
Figure FDA00026058326100000114
and
Figure FDA00026058326100000115
step sizes of convolution kernels of the l' th layer of block convolution layer in the width and height dimensions of the feature map respectively,
Figure FDA00026058326100000116
the operation of rounding up is carried out;
step three, according to the stepThe number m of blocks in each row and column on the input feature map obtained in the second steplAnd nlRandomly sampling to obtain the width of the feature map block
Figure FDA00026058326100000117
And height
Figure FDA00026058326100000118
And is
Figure FDA00026058326100000119
And
Figure FDA00026058326100000120
are all positive integers, and are not limited to the integer,
Figure FDA00026058326100000121
step four, dividing the block width according to the characteristic diagram obtained in the step three
Figure FDA0002605832610000021
And height
Figure FDA0002605832610000022
Feature map x to be input(l)Is divided into ml×nlBlock, obtaining a set of block feature maps
Figure FDA0002605832610000023
Figure FDA0002605832610000024
Step five, adopting the convolution kernel parameter omega in the step one(l)Respectively comparing all obtained in step four
Figure FDA0002605832610000025
Performing convolution to obtain corresponding convolution output characteristic diagram
Figure FDA0002605832610000026
Step six, the convolution output characteristic diagram obtained in the step five is used
Figure FDA0002605832610000027
Splicing according to the original position to obtain the output characteristic diagram of the first convolution layer in the block convolution neural network
Figure FDA0002605832610000028
Step seven, for the L block convolution layers, the operation is carried out according to the steps from one step to six until the output characteristic diagram x of the last L block convolution layer is obtained(L+1)X is to be(L+1)Inputting the data into a full connection layer to obtain the output probability p ∈ R of fine-grained image classificationnAnd n is the number of categories, so that the classification of fine-grained images is realized.
2. The fine-grained image classification method based on the block convolutional neural network as claimed in claim 1, wherein: step eight, adopting cross entropy LCE(t, p) and the real category t optimize the output probability p of the fine-grained image classification:
LCE(t，p)＝-lnpt。
3. the fine-grained image classification method based on the block convolutional neural network as claimed in claim 1, wherein: in the second step, the range of the shrinkage factor on the width dimension and the height dimension of the theoretical receptive field is respectively as follows:
Figure FDA0002605832610000029
4. the fine-grained image classification method based on the block convolutional neural network as claimed in claim 1, wherein: replacing the second step with the sixth step by the following steps:
step A, setting the number m of blocks in each row and column on the output characteristic diagramlAnd nlRandomly sampling to obtain the width of the feature map block
Figure FDA00026058326100000210
And height
Figure FDA00026058326100000211
And is
Figure FDA00026058326100000212
And
Figure FDA00026058326100000213
are all positive integers, and are not limited to the integer,
Figure FDA00026058326100000214
step B, inputting a characteristic diagram x(l)Every other row above
Figure FDA00026058326100000215
Insert into
Figure FDA00026058326100000216
Column all zero column vector, every other row
Figure FDA00026058326100000217
Insert into
Figure FDA00026058326100000218
The rows are all zero row vectors and the row vectors,
Figure FDA00026058326100000219
Figure FDA00026058326100000220
and
Figure FDA00026058326100000221
step sizes of the convolution kernel in the feature width and height dimensions respectively,
Figure FDA0002605832610000031
obtaining a processed feature map for a round-down operation
Figure FDA0002605832610000032
Step C, adopting the convolution kernel parameter omega obtained in the step one(l)To pair
Figure FDA0002605832610000033
Directly performing convolution to obtain convolution output characteristic diagram
Figure FDA0002605832610000034
Step D: c, according to the positions of all zero column vectors and all zero row vectors inserted in the step C, outputting the feature graph by convolution
Figure FDA0002605832610000035
The inserted vector is removed, and the removed column is marked with
Figure FDA0002605832610000036
Figure FDA0002605832610000037
The removed row numbers are
Figure FDA0002605832610000038
Obtaining an output feature map of the first convolutional layer in a partitioned convolutional neural network
Figure FDA0002605832610000039
CN202010738474.1A 2020-07-28 2020-07-28 Fine-grained image classification method based on block convolutional neural network Active CN111860672B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010738474.1A CN111860672B (en) 2020-07-28 2020-07-28 Fine-grained image classification method based on block convolutional neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010738474.1A CN111860672B (en) 2020-07-28 2020-07-28 Fine-grained image classification method based on block convolutional neural network

Publications (2)

Publication Number Publication Date
CN111860672A CN111860672A (en) 2020-10-30
CN111860672B true CN111860672B (en) 2021-03-16

Family

ID=72948450

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010738474.1A Active CN111860672B (en) 2020-07-28 2020-07-28 Fine-grained image classification method based on block convolutional neural network

Country Status (1)

Country Link
CN (1) CN111860672B (en)

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106462549A (en) * 2014-04-09 2017-02-22 尹度普有限公司 Authenticating physical objects using machine learning from microscopic variations
CN106650786A (en) * 2016-11-14 2017-05-10 沈阳工业大学 Image recognition method based on multi-column convolutional neural network fuzzy evaluation
CN107680106A (en) * 2017-10-13 2018-02-09 南京航空航天大学 A kind of conspicuousness object detection method based on Faster R CNN
CN109190622A (en) * 2018-09-11 2019-01-11 深圳辉煌耀强科技有限公司 Epithelial cell categorizing system and method based on strong feature and neural network
CN109191457A (en) * 2018-09-21 2019-01-11 中国人民解放军总医院 A kind of pathological image quality validation recognition methods
CN109344856A (en) * 2018-08-10 2019-02-15 华南理工大学 A kind of off-line signature verification method based on multilayer discriminate feature learning
CN109711448A (en) * 2018-12-19 2019-05-03 华东理工大学 Based on the plant image fine grit classification method for differentiating key field and deep learning
CN110084285A (en) * 2019-04-08 2019-08-02 安徽艾睿思智能科技有限公司 Fish fine grit classification method based on deep learning
CN110110692A (en) * 2019-05-17 2019-08-09 南京大学 A kind of realtime graphic semantic segmentation method based on the full convolutional neural networks of lightweight
US10503978B2 (en) * 2017-07-14 2019-12-10 Nec Corporation Spatio-temporal interaction network for learning object interactions
CN110958187A (en) * 2019-12-17 2020-04-03 电子科技大学 Distributed machine learning parameter-oriented synchronous differential data transmission method

Family Cites Families (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10242036B2 (en) * 2013-08-14 2019-03-26 Ricoh Co., Ltd. Hybrid detection recognition system
US20170061257A1 (en) * 2013-12-16 2017-03-02 Adobe Systems Incorporated Generation of visual pattern classes for visual pattern regonition
CN104331447A (en) * 2014-10-29 2015-02-04 邱桃荣 Cloth color card image retrieval method
CN106951872B (en) * 2017-03-24 2020-11-06 江苏大学 Pedestrian re-identification method based on unsupervised depth model and hierarchical attributes
CN110914829A (en) * 2017-04-07 2020-03-24 英特尔公司 Method and system for image processing using improved convolutional neural network
CN108537283A (en) * 2018-04-13 2018-09-14 厦门美图之家科技有限公司 A kind of image classification method and convolutional neural networks generation method
CN108776807A (en) * 2018-05-18 2018-11-09 复旦大学 It is a kind of based on can the double branch neural networks of skip floor image thickness grain-size classification method
CN109255375A (en) * 2018-08-29 2019-01-22 长春博立电子科技有限公司 Panoramic picture method for checking object based on deep learning
CN109857889B (en) * 2018-12-19 2021-04-09 苏州科达科技股份有限公司 Image retrieval method, device and equipment and readable storage medium
CN109978077B (en) * 2019-04-08 2021-03-12 南京旷云科技有限公司 Visual recognition method, device and system and storage medium
CN111047038A (en) * 2019-11-08 2020-04-21 南昌大学 Neural network compression method using block circulant matrix
CN111178432B (en) * 2019-12-30 2023-06-06 武汉科技大学 Weak supervision fine granularity image classification method of multi-branch neural network model
CN111414954B (en) * 2020-03-17 2022-09-09 重庆邮电大学 Rock image retrieval method and system

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106462549A (en) * 2014-04-09 2017-02-22 尹度普有限公司 Authenticating physical objects using machine learning from microscopic variations
CN106650786A (en) * 2016-11-14 2017-05-10 沈阳工业大学 Image recognition method based on multi-column convolutional neural network fuzzy evaluation
US10503978B2 (en) * 2017-07-14 2019-12-10 Nec Corporation Spatio-temporal interaction network for learning object interactions
CN107680106A (en) * 2017-10-13 2018-02-09 南京航空航天大学 A kind of conspicuousness object detection method based on Faster R CNN
CN109344856A (en) * 2018-08-10 2019-02-15 华南理工大学 A kind of off-line signature verification method based on multilayer discriminate feature learning
CN109190622A (en) * 2018-09-11 2019-01-11 深圳辉煌耀强科技有限公司 Epithelial cell categorizing system and method based on strong feature and neural network
CN109191457A (en) * 2018-09-21 2019-01-11 中国人民解放军总医院 A kind of pathological image quality validation recognition methods
CN109711448A (en) * 2018-12-19 2019-05-03 华东理工大学 Based on the plant image fine grit classification method for differentiating key field and deep learning
CN110084285A (en) * 2019-04-08 2019-08-02 安徽艾睿思智能科技有限公司 Fish fine grit classification method based on deep learning
CN110110692A (en) * 2019-05-17 2019-08-09 南京大学 A kind of realtime graphic semantic segmentation method based on the full convolutional neural networks of lightweight
CN110958187A (en) * 2019-12-17 2020-04-03 电子科技大学 Distributed machine learning parameter-oriented synchronous differential data transmission method

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
A codebook-free and annotation-free approach for fine-grained image categorization;Yao B P等;《Proceedings of the 2012 IEEE Conference on Computer Vision and Pattern Recognition》;20121231;第3466-3473页 *
基于深度卷积特征的细粒度图像分类研究综述;罗建豪等;《自动化学报》;20170217;第43卷(第8期);第1306-1318页 *

Also Published As

Publication number Publication date
CN111860672A (en) 2020-10-30

Similar Documents

Publication Publication Date Title
CN110443842B (en) Depth map prediction method based on visual angle fusion
Wang et al. Detect globally, refine locally: A novel approach to saliency detection
CN109492666B (en) Image recognition model training method and device and storage medium
CN110414394B (en) Facial occlusion face image reconstruction method and model for face occlusion detection
CN111191583B (en) Space target recognition system and method based on convolutional neural network
CN108288035A (en) The human motion recognition method of multichannel image Fusion Features based on deep learning
CN109741341B (en) Image segmentation method based on super-pixel and long-and-short-term memory network
CN107680106A (en) A kind of conspicuousness object detection method based on Faster R CNN
CN112862792B (en) Wheat powdery mildew spore segmentation method for small sample image dataset
CN108710893B (en) Digital image camera source model classification method based on feature fusion
CN108171249B (en) RGBD data-based local descriptor learning method
CN111723915A (en) Pruning method of deep convolutional neural network, computer equipment and application method
CN113705641B (en) Hyperspectral image classification method based on rich context network
CN111401380A (en) RGB-D image semantic segmentation method based on depth feature enhancement and edge optimization
CN114565628B (en) Image segmentation method and system based on boundary perception attention
CN114821058A (en) Image semantic segmentation method and device, electronic equipment and storage medium
CN112101364A (en) Semantic segmentation method based on parameter importance incremental learning
CN111310820A (en) Foundation meteorological cloud chart classification method based on cross validation depth CNN feature integration
Sun et al. Iterative, deep synthetic aperture sonar image segmentation
CN111784699A (en) Method and device for carrying out target segmentation on three-dimensional point cloud data and terminal equipment
CN111860672B (en) Fine-grained image classification method based on block convolutional neural network
CN113361589A (en) Rare or endangered plant leaf identification method based on transfer learning and knowledge distillation
CN111275616B (en) Low-altitude aerial image splicing method and device
CN109949298B (en) Image segmentation quality evaluation method based on cluster learning
CN108805811B (en) Natural image intelligent picture splicing method and system based on non-convex quadratic programming

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