CN108427957A - image classification method and system - Google Patents
image classification method and system Download PDFInfo
- Publication number
- CN108427957A CN108427957A CN201710081054.9A CN201710081054A CN108427957A CN 108427957 A CN108427957 A CN 108427957A CN 201710081054 A CN201710081054 A CN 201710081054A CN 108427957 A CN108427957 A CN 108427957A
- Authority
- CN
- China
- Prior art keywords
- network
- image
- classification
- probability value
- original image
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/254—Fusion techniques of classification results, e.g. of results related to same input data
- G06F18/256—Fusion techniques of classification results, e.g. of results related to same input data of results relating to different input data, e.g. multimodal recognition
Abstract
The present invention relates to a kind of image classification methods, including:Predict that the original image belongs to the probability value of each classification;Judge whether to need to open low resolution network;Low-resolution image is obtained to the original image down-sampling, predicts that the low-resolution image belongs to the probability value of each classification;Judge whether to need to open salient region network;It merges the prediction result of the prediction result of original image network and low resolution network to obtain image category;Conspicuousness is carried out to the original image to detect to obtain salient region image, predicts that the salient region image belongs to the probability value of each classification;It merges the prediction result of the prediction result of original image network, the prediction result of low resolution network and salient region network to obtain image category.The invention further relates to a kind of image classification systems.The present invention can obtain the information of image multi-dimensional degree using multiple dimensioned image information and vision significance, and improve the nicety of grading of image.
Description
Technical field
The present invention relates to a kind of image classification method and systems.
Background technology
Image classification is all widely used in many application fields, for example, target identification, image understanding, based on content
Image retrieval etc..In recent years, breakthrough is obtained in image processing field with deep learning, is schemed using deep learning research
As classification has become a research hotspot.
Currently, many convolutional neural networks require that the input of convolutional layer is image segments (such as the 224* of fixed size
224), because full articulamentum need input identical dimensional feature, and the dimension of feature be by input image segments size,
What convolutional layer and pooling layers of size and step-length determined.
When people watches an object, what closely viewing obtained is more image detail informations, and remote viewing obtains
What is obtained is more profile information, therefore the information of a variety of scales has complementarity to the identification of image.Meanwhile the recipient of image
It is people, the vision attention of people has a great impact to the understanding of image.But utilizing multiple dimensioned image there is presently no a kind of
Information and vision significance obtain the information of image multi-dimensional degree, and carry out the method or system of image classification.
Invention content
In view of this, it is necessary to provide a kind of image classification method and system, multiple dimensioned image information can be utilized
The information of image multi-dimensional degree is obtained with vision significance, and carries out image classification.
The present invention provides a kind of image classification method, and this method comprises the following steps:A. by original image input depth volume
Product network, to predict that the original image belongs to the probability value of each classification, wherein the original image belongs to each classification
Probability value is the prediction result of original image network;B. judge whether to need to open low resolution network:It is low if necessary to open
Resolution ratio network, then enter step c;Otherwise, flow terminates;C. low-resolution image is obtained to the original image down-sampling,
The low-resolution image is inputted into the depth convolutional network, to predict that the low-resolution image belongs to the general of each classification
Rate value, wherein the probability value that the low-resolution image belongs to each classification is the prediction result of low resolution network;D. judge
Whether need to open salient region network:If you do not need to opening salient region network, then e is entered step, flow terminates;
If necessary to open salient region network, then f is entered step;E. by the prediction result of original image network and low resolution net
The prediction result of network merges to obtain image category;F. conspicuousness is carried out to the original image to detect to obtain salient region figure
The salient region image is inputted the depth convolutional network, to predict that the salient region image belongs to each by picture
The probability value of classification, wherein the probability value that the salient region image belongs to each classification is the pre- of salient region network
Survey result;G. by the pre- of the prediction result of original image network, the prediction result of low resolution network and salient region network
Result is surveyed to merge to obtain image category.
Specifically, the depth convolutional network includes:1 input layer, 5 convolutional layers and 3 full articulamentums, wherein:The
1, pooling layers of a Max is respectively met after 2,5 convolutional layers, the last one full articulamentum is output layer.
Specifically, the output layer belongs to the probability value of each classification, the softmax using softmax function predictions
Function is:
Wherein, xiIt is the input of each node in last layer, n is last layer of all node number.
Specifically, the step b is specifically included:It is adopted as the maximum value setting threshold value of the probability value of each dimension prediction
Mode decides whether unlatching next stage dimension network, specially:
Wherein, Pi={ Pi1,Pi2,…,PiCBe i-stage neural network forecast probability value, value range is [0,1], and C is class
Other number,
If the most probable value of original image network output is less than the threshold value T1 of setting, judgement needs to automatically turn on low
Resolution ratio network.
Specifically, the step e is specifically included:
By the probability value weighted blend of original image network and low resolution network:
Pi,j=wiPi+wjPj
Wherein, wiIt is the weight of i-stage network, PiIt is the probability value of i-stage neural network forecast, wjIt is the power of j-th stage network
Weight, PjIt is the probability value of j-th stage neural network forecast, PijIt is the classification for the fusion that i-stage network and j-th stage network weights are mixed to get
Prediction probability value.
The present invention also provides a kind of image classification systems, including original image network module, judgment module, low resolution net
Network module, salient region network module and integrated classification module, wherein:The original image network module is used for original graph
As input depth convolutional network, to predict that the original image belongs to the probability value of each classification, wherein the original image category
In the prediction result that the probability value of each classification is original image network;The judgment module is low for judging whether to need to open
Resolution ratio network;The low resolution network module is used for after opening low resolution network, to the original image down-sampling
Low-resolution image is obtained, the low-resolution image is inputted into the depth convolutional network, to predict the low resolution figure
Probability value as belonging to each classification, wherein the probability value that the low-resolution image belongs to each classification is low resolution net
The prediction result of network;The judgment module is additionally operable to judge whether to need to open salient region network;The integrated classification mould
Block is for merging the prediction result of the prediction result of original image network and low resolution network to obtain image category;It is described aobvious
Work property Local Area Network module is used for after opening salient region network, is detected and is shown to original image progress conspicuousness
The salient region image is inputted the depth convolutional network, to predict the salient region figure by work property area image
Probability value as belonging to each classification, wherein the probability value that the salient region image belongs to each classification is conspicuousness area
The prediction result of domain network;The integrated classification module is additionally operable to the prediction result of original image network, low resolution network
Prediction result and the prediction result of salient region network merge to obtain image category.
Specifically, the convolutional network structure includes:1 input layer, 5 convolutional layers and 3 full articulamentums, wherein:The
1, pooling layers of a Max is respectively met after 2,5 convolutional layers, the last one full articulamentum is output layer.
Specifically, the output layer belongs to the probability value of each classification, the softmax using softmax function predictions
Function is:
Wherein, xiIt is the input of each node in last layer, n is last layer of all node number.
Specifically, the judgment module is specifically used for:
The mode for being adopted as the maximum value setting threshold value of the probability value of each dimension prediction decides whether to open next stage dimension
Network is spent, specially:
Wherein, Pi={ Pi1,Pi2,…,PiCBe i-stage neural network forecast probability value, value range is [0,1], and C is class
Other number,
If the most probable value of original image network output is less than the threshold value T1 of setting, judgement needs to automatically turn on low
Resolution ratio network.
Specifically, the integrated classification module is specifically used for:
By the probability value weighted blend of original image network and low resolution network:
Pi,j=wiPi+wjPj
Wherein, wiIt is the weight of i-stage network, PiIt is the probability value of i-stage neural network forecast, wjIt is the power of j-th stage network
Weight, PjIt is the probability value of j-th stage neural network forecast, PijIt is the classification for the fusion that i-stage network and j-th stage network weights are mixed to get
Prediction probability value..
The present invention can obtain the information of image multi-dimensional degree using multiple dimensioned image information and vision significance, and carry out
Image classification.The low-resolution image and conspicuousness detection information that the present invention is obtained using original-resolution image, down-sampling
The various dimensions feature that image is extracted by depth convolutional network will Bu Tong be led to by the way that rational decision strategy and associative mechanism is arranged
The class probability value in road merges and then improves the nicety of grading of image.
Description of the drawings
Fig. 1 is the flow chart of image classification method of the present invention;
Fig. 2 is the schematic diagram of one embodiment of the invention convolutional network structure;
Fig. 3 is the hardware architecture diagram of image classification system of the present invention.
Specific implementation mode
Below in conjunction with the accompanying drawings and specific embodiment the present invention is described in further detail.
As shown in fig.1, being the operation process chart of image classification method preferred embodiment of the present invention.
Step S1 automatically extracts the original using trained depth convolutional network to original image by being successively abstracted
The high-level semantics features of beginning image recycle last layer of softmax grader in the depth convolutional network, predict the original
Beginning image belongs to the probability value of each classification.In the present embodiment, the depth convolutional network for inputting original image is referred to as original
Image network namely first order network.Wherein, the classification is indicated with K={ 1,2 ..., k }.Specifically:
The convolutional network structure (please referring to Fig.2) that the present embodiment uses, including 1 input layer, 5 convolutional layers and 3 are complete
Articulamentum, wherein:Pooling layers of a Max is respectively met after 1st, 2,5 convolutional layer, the last one full articulamentum is to export
Layer.Neuron number is K, represents the classification number of output, and output layer belongs to the general of each classification using softmax function predictions
Rate value, the softmax functions are:
Wherein, xiIt is the input of each node in last layer, n is last layer of all node number.Softmax functions
Prognostic chart picture belongs to the probability value of each classification.Although the present embodiment uses the convolutional network structure of attached drawing 2, of the invention
Protection domain be not limited to the network, convolutional network structure can also use other depth networks.
Step S2 judges whether to need to open low resolution networkIf necessary to open low resolution network, then enter step
Rapid S3;If you do not need to opening low resolution network, then flow terminates.Specifically:
In the present embodiment, each network becomes a dimension, and each dimension predicts one group of probability value P={ pi, i=
1 ..., K }, wherein piThe input picture for representing the dimension belongs to the probability value of i-th of classification.
The mode that the present embodiment is adopted as the maximum value setting threshold value of the probability value of each dimension prediction decides whether to open
Next stage dimension network.Specially:
Wherein, Pi={ Pi1,Pi2,…,PiCBe i-stage neural network forecast probability value, value range is [0,1], and C is class
Other number.
If the most probable value of original image network output is less than the threshold value T1 of setting, judgement needs to automatically turn on low
Resolution ratio network, enters step S3.
Step S3, by doing the low-resolution image that down-sampling obtains multiple and different scales to the original image, by institute
It states low-resolution image and inputs above-mentioned trained depth convolutional network, automatically extract the high-level semantic of above-mentioned low-resolution image
Feature recycles last layer of softmax grader in the depth convolutional network, predicts that the low-resolution image belongs to
The probability value of each classification.In the present embodiment, the depth convolutional network for inputting low-resolution image is known as low resolution net
Network namely second level network.It specifically includes:
The present embodiment generates low-resolution image using Downsapling method from original image, it is contemplated that calculated performance and meter
Complexity is calculated, the present embodiment uses bilinear interpolation and obtains the low-resolution image.
Wherein, it is extracted described in high-level semantics features and the prediction of the low-resolution image by the low resolution network
The method that low-resolution image belongs to the probability value of each classification is similar with step S1, and which is not described herein again.
Step S4:Judge whether to need to open salient region networkIf you do not need to opening salient region network, then
S5 is entered step, flow terminates;If necessary to open salient region network, then S6 is entered step.Specifically:
If threshold value of the maximum probability not less than setting in original image network and the output valve of the low resolution network integration
When T2, then salient region network need not be opened, S5 is entered step.
If threshold of the maximum probability still less than setting in original image network and the output valve of the low resolution network integration
When value T2, system automatically turns on salient region network, enters step S6.
Specific computational methods are similar with step S2, and which is not described herein again.
Step S5 merges the prediction result of the prediction result of original image network and low resolution network to obtain image class
Not.Specifically:
The present embodiment is by the probability value weighted blend of original image network and low resolution network:
Pi,j=wiPi+wjPj
Wherein, wiIt is the weight of i-stage network, PiIt is the probability value of i-stage neural network forecast, wjIt is the power of j-th stage network
Weight, PjIt is the probability value of j-th stage neural network forecast, PijIt is the classification for the fusion that i-stage network and j-th stage network weights are mixed to get
Prediction probability value.
Step S6 carries out conspicuousness to original image and detects to obtain the salient region namely salient region figure of image
Picture, and the salient region image is normalized, the salient region image is inputted into above-mentioned trained depth
Convolutional network extracts the high-level semantics features of the salient region image, and utilizes last in the depth convolutional network
One layer of softmax grader predicts that the salient region image belongs to the probability value of each classification.In the present embodiment, will
The depth convolutional network of input salient region image is known as salient region network namely third level network.Specifically:
The present embodiment detects salient region using conspicuousness detection method from original image, selects and is based on histogram pair
Conspicuousness detection method than degree generates saliency region mask, and this method divides the image into region first, then passes through
The global contrast angle value of zoning is that significance value is distributed in each region, eventually forms the conspicuousness based on region contrast
Figure.The mask that the present embodiment detects conspicuousness, chooses the minimum rectangle encirclement frame of foreground in mask, then scales it
The picture size that the salient region network needs.
Wherein, high-level semantics features and the prediction of the salient region image are extracted by the salient region network
The method that the salient region image belongs to the probability value of each classification is similar with step S1, and which is not described herein again.
Step S7, by the prediction result of original image network, the prediction result of low resolution network and salient region net
The prediction result of network merges to obtain image category.
Specific fusion method is similar with step S5, and which is not described herein again.
In order to briefly explain the main points of invention, the present embodiment only lists three dimensions, but present invention model to be protected
It encloses and is not limited to three dimensions, can be merged with the information of multiple dimensions.
As shown in fig.3, being the hardware architecture diagram of image classification system 10 of the present invention.The system includes original image network
Module 101, judgment module 102, low resolution network module 103, salient region network module 104 and integrated classification module
105。
The original image network module 101 be used for utilize trained depth convolutional network, to original image by by
Layer is abstract, automatically extracts the high-level semantics features of the original image, recycles last layer in the depth convolutional network
Softmax graders predict that the original image belongs to the probability value of each classification.In the present embodiment, original graph will be inputted
The depth convolutional network of picture is known as original image network.Wherein, the classification is indicated with K={ 1,2 ..., k }.Specifically:
The convolutional network structure (please referring to Fig.2) that the present embodiment uses, including 1 input layer, 5 convolutional layers and 3 are complete
Articulamentum, wherein:Pooling layers of a Max is respectively met after 1st, 2,5 convolutional layer, the last one full articulamentum is to export
Layer.Neuron number is K, represents the classification number of output, and output layer belongs to the general of each classification using softmax function predictions
Rate value, the softmax functions are:
Wherein, xiIt is the input of each node in last layer, n is last layer of all node number.Softmax functions
Prognostic chart picture belongs to the probability value of each classification.Although the present embodiment uses the convolutional network structure of attached drawing 2, of the invention
Protection domain be not limited to the network, convolutional network structure can also use other depth networks.
The judgment module 102 needs to open low resolution network for judging whether.Specifically:
In the present embodiment, each network becomes a dimension, and each dimension predicts one group of probability value P={ pi, i=
1 ..., K }, wherein piThe input picture for representing the dimension belongs to the probability value of i-th of classification.
The mode that the present embodiment is adopted as the maximum value setting threshold value of the probability value of each dimension prediction decides whether to open
Next stage dimension network.Specially:
Wherein, Pi={ Pi1,Pi2,…,PiCBe i-stage neural network forecast probability value, value range is [0,1], and C is class
Other number.
If the most probable value of original image network output is less than the threshold value T1 of setting, the judgement of judgment module 102 needs
Low resolution network is automatically turned on, otherwise flow terminates.
The low resolution network module 103 is used for after opening low resolution network, by being done to the original image
Down-sampling obtains the low-resolution image of multiple and different scales, and the above-mentioned trained depth of low-resolution image input is rolled up
Product network, automatically extracts the high-level semantics features of above-mentioned low-resolution image, recycles last in the depth convolutional network
One layer of softmax grader predicts that the low-resolution image belongs to the probability value of each classification.It in the present embodiment, will be defeated
The depth convolutional network for entering low-resolution image is known as low resolution network.It specifically includes:
The low resolution network module 103 generates low-resolution image using Downsapling method from original image, examines
Consider calculated performance and computation complexity, the present embodiment uses bilinear interpolation and obtains the low-resolution image.
Wherein, it is extracted described in high-level semantics features and the prediction of the low-resolution image by the low resolution network
The method that low-resolution image belongs to the probability value of each classification is similar with original image network module 101, no longer superfluous here
It states.
The judgment module 102 is additionally operable to judge whether to need to open salient region network.Specifically:
If threshold value of the maximum probability not less than setting in original image network and the output valve of the low resolution network integration
When T2, then salient region network need not be opened.
If threshold of the maximum probability still less than setting in original image network and the output valve of the low resolution network integration
When value T2, system automatically turns on salient region network.
Specifically computational methods judge whether to need unlatching low resolution network similar with aforementioned, and which is not described herein again.
The integrated classification module 105 is used for the prediction knot of the prediction result of original image network and low resolution network
Fruit merges to obtain image category.Specifically:
The integrated classification module 105 is by the probability value weighted blend of original image network and low resolution network:
Pi,j=wiPi+wjPj
Wherein, wiIt is the weight of i-stage network, PiIt is the probability value of i-stage neural network forecast, wjIt is the power of j-th stage network
Weight, PjIt is the probability value of j-th stage neural network forecast, PijIt is the classification for the fusion that i-stage network and j-th stage network weights are mixed to get
Prediction probability value.
The salient region network module 104 is used for after opening salient region network, is shown to original image
Work property detects to obtain the salient region of image namely salient region image, and returns to the salient region image
One changes, and the salient region image is inputted above-mentioned trained depth convolutional network, extracts the salient region image
High-level semantics features, and using last layer of softmax grader in the depth convolutional network, predict the conspicuousness
Area image belongs to the probability value of each classification.In the present embodiment, the depth convolutional network of salient region image will be inputted
Referred to as salient region network.Namely:
The present embodiment detects salient region using conspicuousness detection method from original image, selects and is based on histogram pair
Conspicuousness detection method than degree generates saliency region mask, and this method divides the image into region first, then passes through
The global contrast angle value of zoning is that significance value is distributed in each region, eventually forms the conspicuousness based on region contrast
Figure.The mask that the present embodiment detects conspicuousness, chooses the minimum rectangle encirclement frame of foreground in mask, then scales it
The picture size that the salient region network needs.
Wherein, high-level semantics features and the prediction of the salient region image are extracted by the salient region network
The method that the salient region image belongs to the probability value of each classification is similar with original image network module 101, here
It repeats no more.
The integrated classification module 105 is additionally operable to the prediction of the prediction result of original image network, low resolution network
As a result and the prediction result of salient region network merges to obtain image category.
Specific fusion method is with aforementioned by the prediction result of original image network and the prediction result of low resolution network
Merge similar, which is not described herein again.
In order to briefly explain the main points of invention, the present embodiment only lists three dimensions, but present invention model to be protected
It encloses and is not limited to three dimensions, can be merged with the information of multiple dimensions.
The present invention is using original image, low-resolution image, salient region image respectively with trained depth convolution
Network carries out class prediction, and each network becomes a dimension, by the way that decision strategy and syncretizing mechanism is arranged, is turned on or off
Low resolution network or salient region network, finally merge the prediction result of multiple and different networks to obtain image category.
Although the present invention is described with reference to current better embodiment, those skilled in the art should be able to manage
Solution, for above-mentioned better embodiment only for illustrating the present invention, protection domain not for the purpose of limiting the invention is any in the present invention
Spirit and spirit within, any modification, equivalence replacement, improvement for being done etc., should be included in the present invention right protect
Within the scope of shield.
Claims (10)
1. a kind of image classification method, which is characterized in that the method comprising the steps of:
A. original image is inputted into depth convolutional network, to predict that the original image belongs to the probability value of each classification, wherein
The probability value that the original image belongs to each classification is the prediction result of original image network;
B. judge whether to need to open low resolution network:If necessary to open low resolution network, then c is entered step;Otherwise,
Flow terminates;
C. low-resolution image is obtained to the original image down-sampling, the low-resolution image is inputted into the depth convolution
Network, to predict that the low-resolution image belongs to the probability value of each classification, wherein the low-resolution image belongs to each
The probability value of classification is the prediction result of low resolution network;
D. judge whether to need to open salient region network:If you do not need to opening salient region network, then enter step
E, flow terminate;If necessary to open salient region network, then f is entered step;
E. it merges the prediction result of the prediction result of original image network and low resolution network to obtain image category;
F. it carries out conspicuousness to the original image to detect to obtain salient region image, the salient region image is inputted
The depth convolutional network, to predict that the salient region image belongs to the probability value of each classification, wherein the conspicuousness
The probability value that area image belongs to each classification is the prediction result of salient region network;
G. by the prediction knot of the prediction result of original image network, the prediction result of low resolution network and salient region network
Fruit merges to obtain image category.
2. the method as described in claim 1, which is characterized in that the depth convolutional network includes:1 input layer, 5 convolution
Layer and 3 full articulamentums, wherein:Pooling layers of a Max is respectively met after 1st, 2,5 convolutional layer, the last one full articulamentum
For output layer.
3. method as claimed in claim 2, which is characterized in that the output layer belongs to each using softmax function predictions
The probability value of classification, the softmax functions are:
Wherein, xiIt is the input of each node in last layer, n is last layer of all node number.
4. method as claimed in claim 3, which is characterized in that the step b is specifically included:
The mode for being adopted as the maximum value setting threshold value of the probability value of each dimension prediction decides whether to open next stage dimension net
Network, specially:
Wherein, Pi={ Pi1,Pi2,…,PiCBe i-stage neural network forecast probability value, value range is [0,1], and C is classification
Number,
If the most probable value of original image network output is less than the threshold value T1 of setting, judgement needs to automatically turn on low resolution
Rate network.
5. method as claimed in claim 4, which is characterized in that the step e is specifically included:
By the probability value weighted blend of original image network and low resolution network:
Pi,j=wiPi+wjPj
Wherein, wiIt is the weight of i-stage network, PiIt is the probability value of i-stage neural network forecast, wjIt is the weight of j-th stage network, PjIt is
The probability value of j-th stage neural network forecast, PijBe the fusion that i-stage network and j-th stage network weights are mixed to get class prediction it is general
Rate value.
6. a kind of image classification system, which is characterized in that the system includes original image network module, judgment module, low resolution
Rate network module, salient region network module and integrated classification module, wherein:
The original image network module is used to original image inputting depth convolutional network, to predict that the original image belongs to
The probability value of each classification, wherein the probability value that the original image belongs to each classification is the prediction knot of original image network
Fruit;
The judgment module needs to open low resolution network for judging whether;
The low resolution network module is used for after opening low resolution network, and low point is obtained to the original image down-sampling
The low-resolution image is inputted the depth convolutional network, to predict that the low-resolution image belongs to every by resolution image
The probability value of a classification, wherein the probability value that the low-resolution image belongs to each classification is the prediction of low resolution network
As a result;
The judgment module is additionally operable to judge whether to need to open salient region network;
The integrated classification module is used to merge the prediction result of original image network and the prediction result of low resolution network
Obtain image category;
The salient region network module is used for after opening salient region network, and conspicuousness is carried out to the original image
Detection obtains salient region image, and the salient region image is inputted the depth convolutional network, described aobvious to predict
Work property area image belongs to the probability value of each classification, wherein the salient region image belongs to the probability value of each classification
For the prediction result of salient region network;
The integrated classification module is additionally operable to the prediction result of original image network, the prediction result of low resolution network and shows
The prediction result of work property Local Area Network merges to obtain image category.
7. system as claimed in claim 6, which is characterized in that the convolutional network structure includes:1 input layer, 5 convolution
Layer and 3 full articulamentums, wherein:Pooling layers of a Max is respectively met after 1st, 2,5 convolutional layer, the last one full articulamentum
For output layer.
8. system as claimed in claim 7, which is characterized in that the output layer belongs to each using softmax function predictions
The probability value of classification, the softmax functions are:
Wherein, xiIt is the input of each node in last layer, n is last layer of all node number.
9. system as claimed in claim 8, which is characterized in that the judgment module is specifically used for:
The mode for being adopted as the maximum value setting threshold value of the probability value of each dimension prediction decides whether to open next stage dimension net
Network, specially:
Wherein, Pi={ Pi1,Pi2,…,PiCBe i-stage neural network forecast probability value, value range is [0,1], and C is classification
Number,
If the most probable value of original image network output is less than the threshold value T1 of setting, judgement needs to automatically turn on low resolution
Rate network.
10. method as claimed in claim 9, which is characterized in that the integrated classification module is specifically used for:
By the probability value weighted blend of original image network and low resolution network:
Pi,j=wiPi+wjPj
Wherein, wiIt is the weight of i-stage network, PiIt is the probability value of i-stage neural network forecast, wjIt is the weight of j-th stage network, PjIt is
The probability value of j-th stage neural network forecast, PijBe the fusion that i-stage network and j-th stage network weights are mixed to get class prediction it is general
Rate value.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710081054.9A CN108427957B (en) | 2017-02-15 | 2017-02-15 | Image classification method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710081054.9A CN108427957B (en) | 2017-02-15 | 2017-02-15 | Image classification method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108427957A true CN108427957A (en) | 2018-08-21 |
CN108427957B CN108427957B (en) | 2021-12-21 |
Family
ID=63155279
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710081054.9A Active CN108427957B (en) | 2017-02-15 | 2017-02-15 | Image classification method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108427957B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109635093A (en) * | 2018-12-17 | 2019-04-16 | 北京百度网讯科技有限公司 | Method and apparatus for generating revert statement |
US11409999B2 (en) | 2020-04-21 | 2022-08-09 | Industrial Technology Research Institute | Method of labelling features for image recognition and apparatus thereof |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105320945A (en) * | 2015-10-30 | 2016-02-10 | 小米科技有限责任公司 | Image classification method and apparatus |
CN105512683A (en) * | 2015-12-08 | 2016-04-20 | 浙江宇视科技有限公司 | Target positioning method and device based on convolution neural network |
CN105551036A (en) * | 2015-12-10 | 2016-05-04 | 中国科学院深圳先进技术研究院 | Training method and device for deep learning network |
US20160140424A1 (en) * | 2014-11-13 | 2016-05-19 | Nec Laboratories America, Inc. | Object-centric Fine-grained Image Classification |
CN105631466A (en) * | 2015-12-21 | 2016-06-01 | 中国科学院深圳先进技术研究院 | Method and device for image classification |
CN106372648A (en) * | 2016-10-20 | 2017-02-01 | 中国海洋大学 | Multi-feature-fusion-convolutional-neural-network-based plankton image classification method |
-
2017
- 2017-02-15 CN CN201710081054.9A patent/CN108427957B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20160140424A1 (en) * | 2014-11-13 | 2016-05-19 | Nec Laboratories America, Inc. | Object-centric Fine-grained Image Classification |
CN105320945A (en) * | 2015-10-30 | 2016-02-10 | 小米科技有限责任公司 | Image classification method and apparatus |
CN105512683A (en) * | 2015-12-08 | 2016-04-20 | 浙江宇视科技有限公司 | Target positioning method and device based on convolution neural network |
CN105551036A (en) * | 2015-12-10 | 2016-05-04 | 中国科学院深圳先进技术研究院 | Training method and device for deep learning network |
CN105631466A (en) * | 2015-12-21 | 2016-06-01 | 中国科学院深圳先进技术研究院 | Method and device for image classification |
CN106372648A (en) * | 2016-10-20 | 2017-02-01 | 中国海洋大学 | Multi-feature-fusion-convolutional-neural-network-based plankton image classification method |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109635093A (en) * | 2018-12-17 | 2019-04-16 | 北京百度网讯科技有限公司 | Method and apparatus for generating revert statement |
US11409999B2 (en) | 2020-04-21 | 2022-08-09 | Industrial Technology Research Institute | Method of labelling features for image recognition and apparatus thereof |
Also Published As
Publication number | Publication date |
---|---|
CN108427957B (en) | 2021-12-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Zhao et al. | Deep temporal convolutional networks for short-term traffic flow forecasting | |
US11367272B2 (en) | Target detection method, apparatus, and system | |
CN109815886B (en) | Pedestrian and vehicle detection method and system based on improved YOLOv3 | |
Yuan et al. | Gated CNN: Integrating multi-scale feature layers for object detection | |
CN109635694B (en) | Pedestrian detection method, device and equipment and computer readable storage medium | |
CN110503112A (en) | A kind of small target deteection of Enhanced feature study and recognition methods | |
CN113572742B (en) | Network intrusion detection method based on deep learning | |
CN113379771B (en) | Hierarchical human body analysis semantic segmentation method with edge constraint | |
CN110991444B (en) | License plate recognition method and device for complex scene | |
CN111738258A (en) | Pointer instrument reading identification method based on robot inspection | |
CN108986453A (en) | A kind of traffic movement prediction method based on contextual information, system and device | |
CN111339849A (en) | Pedestrian re-identification method integrating pedestrian attributes | |
CN106682092A (en) | Target retrieval method and terminal | |
CN113362491A (en) | Vehicle track prediction and driving behavior analysis method | |
CN107506792A (en) | A kind of semi-supervised notable method for checking object | |
CN114612476B (en) | Image tampering detection method based on full-resolution hybrid attention mechanism | |
CN111400572A (en) | Content safety monitoring system and method for realizing image feature recognition based on convolutional neural network | |
CN110751195A (en) | Fine-grained image classification method based on improved YOLOv3 | |
Xu et al. | BANet: A balanced atrous net improved from SSD for autonomous driving in smart transportation | |
CN108427957A (en) | image classification method and system | |
Pan et al. | A network traffic classification method based on graph convolution and lstm | |
Moghaddam et al. | Jointly human semantic parsing and attribute recognition with feature pyramid structure in EfficientNets | |
CN113298186A (en) | Network abnormal flow detection method for confluent flow model confrontation generation network and clustering algorithm | |
CN111783688B (en) | Remote sensing image scene classification method based on convolutional neural network | |
Kumar et al. | Robust Vehicle Detection Based on Improved You Look Only Once. |
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 |