CN107122701A - A kind of traffic route sign based on saliency and deep learning - Google Patents
A kind of traffic route sign based on saliency and deep learning Download PDFInfo
- Publication number
- CN107122701A CN107122701A CN201710122736.XA CN201710122736A CN107122701A CN 107122701 A CN107122701 A CN 107122701A CN 201710122736 A CN201710122736 A CN 201710122736A CN 107122701 A CN107122701 A CN 107122701A
- Authority
- CN
- China
- Prior art keywords
- msub
- mrow
- image
- region
- traffic route
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
- G06V20/582—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of traffic signs
-
- 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
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Neurology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of traffic route sign based on saliency and deep learning.Methods described generates a network (Alexnet) by convolutional neural networks (CNN) first, then by obtaining traffic route mark picture from ImageNet data sets, training picture is normalized after pretreatment and inputs network, AlexNet networks are then trained.The Alexnet convolutional neural networks trained will be inputted after test pictures progress RC conspicuousness extractions again, carry out the testing classification of traffic route mark.Present invention incorporates the advantage of saliency high efficiency extraction marking area and deep learning in picture recognition, the purpose that exact classification is carried out to traffic route mark picture is reached.
Description
Technical field
The invention belongs to image identification technical field, more particularly to a kind of traffic road based on convolutional neural networks model
The image-recognizing method of road sign will.
Background technology
With expanding economy, Modern Traffic is very flourishing, and road traffic still occupies leading position.But, hand over
Logical safety and traffic jam are therewith into increasingly serious social concern, while also result in surprising economic loss.Add
The problems such as environmental pollution and the energy, the solution for allowing for road traffic problem must not be without recourse to intellectual technology.Intelligent transportation system
This research field of uniting just is arisen at the historic moment, and is developed rapidly.
Traffic sign recognition is one of still unsolved problem in intelligent transportation system research field, be also difficulty compared with
One of big outdoor scene figure identification problem.Although the research work of traffic sign recognition has been achieved for certain achievement,
But also there is many problems and defect.Mainly have：
1st, the comparison for laying particular stress on theoretical research is more, and application-oriented research is lacked.The proposition of many theoretical and methods is all to set up
On the basis of standard drawing or certain local circumstance, without in view of the demand in more practical applications.
2nd, processing method is relatively simple, and intelligent method is few, and the combination of intelligent method and other method is seldom.Use at present
Processing method has also mainly used the intelligent methods such as artificial neural network and genetic algorithms, and many methods are all the improvement to conventional method
Or application, but intelligent method can be seldom combined with other method.
3rd, most of experimental subjects are all based on standard drawing, and the research for realistic picture is less.Differentiate link with gray scale
Figure is many of object, lacking by object of cromogram.
The content of the invention
It is a primary object of the present invention to overcome the shortcoming and deficiency of prior art there is provided one kind based on saliency and
The traffic route sign of deep learning, the present invention using saliency extraction algorithm to detection image at
Manage, efficiently this salient region of extraction traffic route mark, it is suppressed that the interference in non-significant region, can more protrude the complicated back of the body
Well-marked target under scape, improves treatment effeciency and accuracy rate, with more preferable robustness, is more suitable for actual intelligent transportation road
In the system of road.
In order to achieve the above object, the present invention uses following technical scheme：
A kind of course of work based on saliency and the traffic route sign of the deep learning present invention, bag
Include following step：
S1, download training dataset；
S2, pretreatment, are pre-processed to the image in data set in step S1；
S3, training convolutional neural networks model；
S4, conspicuousness are extracted；Region contrast method is employed, the road signs in test pictures are extracted；
S5, the road signs picture extracted in step S4 is input to the convolutional Neural net trained in step S3
In network model, the classification of traffic route sign image is gone out by convolutional neural networks Model Identification.
It is preferred that, download training dataset in step S1, data set from 2015 national fuzzy image processing contest with
Traffic Sign Recognition training set in content analysis contest, the traffic mark being related to is divided into Warning Sign, Prohibition Sign, instruction
Mark, way-finding sign, tourist district's mark, 7 major classes of road construction secure ID and accessory ID, altogether comprising 72 kinds of traffic marks
Know, relevant criterion of the mark title from country and the Ministry of Public Security, such as T-junction.The quality identified in image includes clear, mould
Paste, block, 5 types such as shade and inclination.
It is preferred that, step S2 pretreatments are as follows：
Each image, scales the images to 256 × 256 sizes, is then intercepted at image upper, middle and lower or so 55 big
It is small be 224 × 224 standard size image, and the classification of standard size image corresponding thereto is preserved in pairs, " a standard
Sized image-classification " is to being used as a data.
It is preferred that, step S3 convolutional neural networks workflows are as follows, and input layer is made up of 32 × 32 sensing nodes, is connect
Receive original image A, then, calculation process between convolution and sub-sample alternately, it is described in detail below：
First hidden layer carries out convolution, and it is made up of 8 Feature Mappings, and each Feature Mapping is by 28 × 28 neural tuples
Into each neuron specifies the receiving of one 5 × 5；
Second hidden layer realizes sub-sample and local average, and it is equally made up of 8 Feature Mappings, but each of which feature is reflected
Penetrate and be made up of 14 × 14 neurons.Each neuron has the acceptance region of one 2 × 2, and one can train biasing and one
Sigmoid activation primitives；The operating point of coefficient and biasing control neuron can be trained；
3rd hidden layer carries out second of convolution, and it is made up of 20 Feature Mappings, and each Feature Mapping is by 10 × 10
Neuron is constituted；Each neuron in the hidden layer may have the cynapse being connected with the several Feature Mappings of next hidden layer
Connection, it to first similar mode of convolutional layer to operate；
4th hidden layer carries out second of sub-sample and local average computation；It is made up of 20 Feature Mappings, but each
Feature Mapping is made up of 5 × 5 neurons, and it to first time similar mode of sampling to operate；
5th hidden layer realizes the final stage of convolution, and it is made up of 120 neurons, and each neuron specifies one 5
× 5 acceptance region；
It is finally a full articulamentum, obtains output vector；
Connection alternating of the successive computation layer between convolution and sampling, we obtain the effect of one " double pointed tower ",
It is exactly in each convolution or sampling layer, as spatial resolution declines, the Feature Mapping quantity increase compared with corresponding preceding layer.
The thought that sub-sample is carried out after convolution is to be followed by " complicated " cell by " simple " cell in animal vision system
Idea inspiration and produce.
It is preferred that, step S4 carries out conspicuousness processing to test pictures, employs region contrast method, i.e. RC methods,
Spatial relationship is incorporated into area grade contrast and calculates central；It is as follows that conspicuousness extraction step is carried out with RC methods：
S4.1, original image A will be inputted first be divided into color contrast on several target areas, then zoning grade
Degree；Input picture is divided into several regions using a kind of image partition method based on figure, is that color is set up in each region
Histogram, for a region rk, its conspicuousness is calculated by measuring it with the color contrast in other regions in image,
It is as follows：
Wherein w (ri) it is region riWeights, Dr(rk,ri) it is two interregional color measurements；riIn number of pixels w
(ri) emphasize and the color contrast in bigger region；Region r1And r2Between color distance be defined as foloows：
In order to preferably reflect a difference between color and primary color, we appear in this using this color
Frequency in region is used as the weights of this color, wherein f (ck, i) it is i-th of color ck,iIn k-th of region rkIn it is all
NkThe frequency of occurrences in individual color.
S4.2, the region contrast for being re-introduced on the basis of step S4.1 spatial weighting；One is introduced comprising space to believe
The weight of breath, the setting of weights is defined according to regional space distance, the target area of space length farther out be endowed compared with
Small weights, the nearer target area of space length is endowed larger weights；Increase the influence of immediate area with this and subtract
Few influence compared with far region；For any region rk, the conspicuousness based on spatial weighting region contrast is defined as follows：
Wherein, Ds(rk,ri) it is region rkAnd riBetween space length, σsControl the intensity of space weights；Larger σs
Value can more reduce the influence of space weights so that region farther out is more conducive to obtain after the conspicuousness of current region, processing
Gray level image B；
S4.3, the salient region for extracting test image, obtain including the gray level image B of traffic route identified areas,
Next binaryzation is carried out to gray level image B and obtains bianry image C.
Holes filling in S4.4, the binary image C for obtaining step S4.3, obtains image D, then using image D as
Mask, retains the part that D correspondence positions are 1, other are set to 0, obtain including traffic route mark region in original image A
Saliency maps picture.
It is preferred that, binary image uses global threshold algorithm in step S4.3, and Otsu methods are carried out at binaryzation to image
Reason；Otsu is the statistical property based on entire image, realizes the automatic selection of threshold value；If image has M gray value, span
In 0~M-1, gray value t is chosen within this range, two groups of G are divided the image into0And G1, G0Comprising pixel gray value 0~
t,G1Gray value in t+1~M-1, total number of image pixels, n are represented with NiRepresent number of the gray value for i pixel, then it is each
The Probability p that individual gray value i occursiFor：
G0Class occur probability and average be：
Wherein w0For G0The probability that class occurs, u0For G0The average of class；
G1Class occur probability and average be：
Wherein w1For G1The probability that class occurs, u1For G1The average of class；
Then inter-class variance δ (t)2=w0w1(u0-u1)2
Global threshold T is the value for the t for making inter-class variance maximum, is：
T=argmax δ (t), t ∈ [0, M-1]
Entire image is divided into foreground and background two parts by threshold value T, the gray value of every pixel is more than this threshold value
225 are set as, 0 is set as less than this threshold value；Gray level image is processed into only black and white binary image, finally given
Binary image C.
It is preferred that, step S5 is according to training model measurement image recognition effect：Included after conspicuousness in S4 is extracted
Test experiments are carried out in the convolutional neural networks trained in the picture input S3 for having road signs region.
The present invention compared with prior art, has the following advantages that and beneficial effect：
1st, the present invention is handled detection image using saliency extraction algorithm, efficiently extracts traffic route mark
This salient region of will, it is suppressed that the interference in non-significant region, can more protrude the well-marked target under complex background, raising processing
Efficiency and accuracy rate, with more preferable robustness, are more suitable in actual intelligent transportation roadnet.
2nd, the present invention uses convolutional neural networks model, and the special construction that convolutional neural networks are shared with its local weight exists
There is the superiority of uniqueness in terms of speech recognition and image procossing, it is laid out closer to actual biological neural network, weights
The shared complexity for reducing network, the image of particularly many dimensional input vectors can directly input network this feature and avoid
Feature extraction and the complexity of assorting process data reconstruction.
3rd, the present invention breaks that art methods are single, the few limitation of intelligent method, with reference to image processing method and depth
Learning method.
Brief description of the drawings
Fig. 1 is particular flow sheet of the embodiment of the present invention；
Fig. 2 is the concrete structure of the convolutional neural networks of the embodiment of the present invention；
Fig. 3 extracts the flow chart of traffic route mark for the conspicuousness of the embodiment of the present invention.
Embodiment
With reference to embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention are not limited
In this.
Embodiment
It is as shown in Figure 1 flow chart of the method for the present invention, including saliency is extracted, normalized, training convolutional
Four parts such as neural network model and test convolutional network model.Comprise the following steps that：
1st, training dataset is downloaded, data set is in national fuzzy image processing contest and content analysis contest in 2015
Traffic Sign Recognition training set, the traffic mark being related to be divided into Warning Sign, Prohibition Sign, sign, way-finding sign,
Tourist district's mark, 7 major classes of road construction secure ID and accessory ID, altogether comprising 72 kinds of traffic marks, mark title is come
From the relevant criterion of country and the Ministry of Public Security, such as T-junction.That the quality identified in image includes is clear, obscure, block, shade and
5 types such as inclination.
2nd, pre-process, each image scales the images to 256 × 256 sizes, then at image upper, middle and lower or so 5
Interception it is 5 big it is small be 224 × 224 standard size image, and the classification of standard size image corresponding thereto is preserved in pairs,
One " standard size image-classification " is to being used as a data.
3rd, training convolutional neural networks model, the input layer of convolutional neural networks is made up of 32 × 32 sensing nodes, is connect
Original image A is received, then, calculation process alternately, is illustrated in figure 2 the convolution of the present embodiment between convolution and sub-sample
The concrete structure of neutral net, concrete structure is as described below：
First hidden layer carries out convolution, and it is made up of 8 Feature Mappings, and each Feature Mapping is by 28 × 28 neural tuples
Into each neuron specifies the receiving of one 5 × 5；
Second hidden layer realizes sub-sample and local average, and it is equally made up of 8 Feature Mappings, but each of which feature is reflected
Penetrate and be made up of 14 × 14 neurons.Each neuron has the acceptance region of one 2 × 2, and one can train biasing and one
Sigmoid activation primitives；The operating point of coefficient and biasing control neuron can be trained；
3rd hidden layer carries out second of convolution, and it is made up of 20 Feature Mappings, and each Feature Mapping is by 10 × 10
Neuron is constituted；Each neuron in the hidden layer may have the cynapse being connected with the several Feature Mappings of next hidden layer
Connection, it to first similar mode of convolutional layer to operate；
4th hidden layer carries out second of sub-sample and local average computation；It is made up of 20 Feature Mappings, but each
Feature Mapping is made up of 5 × 5 neurons, and it to first time similar mode of sampling to operate；
5th hidden layer realizes the final stage of convolution, and it is made up of 120 neurons, and each neuron specifies one 5
× 5 acceptance region；
It is finally a full articulamentum, obtains output vector；
Connection alternating of the successive computation layer between convolution and sampling, we obtain the effect of one " double pointed tower ",
It is exactly in each convolution or sampling layer, as spatial resolution declines, the Feature Mapping quantity increase compared with corresponding preceding layer.
The thought that sub-sample is carried out after convolution is to be followed by " complicated " cell by " simple " cell in animal vision system
Idea inspiration and produce.
4th, conspicuousness is extracted, and refer to Analysis of Contrast method：Region contrast is RC methods, and spatial relationship is incorporated into
Among area grade contrast is calculated.The conspicuousness for being illustrated in figure 3 the present embodiment extracts the flow chart of traffic route mark, shows
It is as follows that work property extracts detailed process：
4.1st, input picture A is divided into the color contrast on several target areas, then zoning grade first.Make
Input picture is divided into several regions with a kind of image partition method based on figure, is that color histogram is set up in each region
Figure, for a region rk, its conspicuousness is calculated by measuring it with the color contrast in other regions in image, such as
Under：
Wherein S (rk) be image r in region rkSaliency value, Dr(rk,ri) it is region r in the Lab of spacekWith region ri
Two interregional color measurements, w (ri) it is riIn number of pixels, to emphasize and the color contrast in bigger region.
Region r1And r2Between color distance be defined as foloows：
In order to preferably reflect a difference between color and primary color, we appear in this using this color
Frequency in region is used as the weights of this color, wherein f (ck, i) it is i-th of color ck,iIn k-th of region rkIn it is all
NkThe frequency of occurrences in individual color, is represented to improve region storage and the computational efficiency present invention using a sparse histograms
Method is efficiently stored and calculated.
4.2nd, the region contrast of spatial weighting is re-introduced on the basis of step 4.1.We introduce one and include space
The weight of information, the setting of weights is defined according to regional space distance, and the target area of space length farther out is endowed
Less weights, the nearer target area of space length is endowed larger weights.Increase the influence of immediate area with this simultaneously
Reduce the influence compared with far region.For any region rk, the conspicuousness based on spatial weighting region contrast is defined as follows：
Wherein, Ds(rk,ri) it is region rkAnd riBetween space length, the space length between two regions is defined as
Euclidean distance between the center of gravity in respective region.σsControl the intensity of space weights.Larger σsValue can more reduce space
The influence of weights so that region farther out is more conducive to the conspicuousness of current region.We use in the present inventionGray level image B is obtained after processing.
4.3rd, the salient region of test image is extracted, obtains including the gray level image B of traffic route identified areas, connects
Get off and bianry image C is obtained to gray level image B progress binaryzation.It is specific as follows：
Binary image uses global threshold algorithm, and Otsu methods carry out binary conversion treatment to image；Otsu is to be based on view picture
The statistical property of image, realizes the automatic selection of threshold value；If image has M gray value, span is in 0~M-1, in this scope
Interior selection gray value t, divides the image into two groups of G0And G1, G0Comprising pixel gray value in 0~t, G1Gray value in t+1
~M-1, total number of image pixels, n are represented with NiNumber of the gray value for i pixel is represented, then each gray value i appearance is general
Rate piFor：
G0Class occur probability and average be：
Wherein w0For G0The probability that class occurs, u0For G0The average of class；
G1Class occur probability and average be：
Wherein w1For G1The probability that class occurs, u1For G1The average of class；
Then inter-class variance δ (t)2=w0w1(u0-u1)2
Global threshold T is the value for the t for making inter-class variance maximum, is：
T=argmax δ (t), t ∈ [0, M-1]
Entire image is divided into foreground and background two parts by threshold value T, the gray value of every pixel is more than this threshold value
225 are set as, 0 is set as less than this threshold value；Gray level image is processed into only black and white binary image, finally given
Binary image C.
4.4th, the holes filling in the binary image C for obtaining step 4.3, obtains image D, then using image D as covering
Mould, retains the part that D correspondence positions are 1, other are set to 0, obtain including traffic route mark region in original image A
Saliency maps picture.
5th, basis trains model measurement image recognition effect：Include road after conspicuousness in step 4 is extracted to hand over
Test experiments are carried out in the convolutional neural networks trained in the picture input step 3 of logical mark region.
Above-described embodiment is preferably embodiment, but embodiments of the present invention are not by above-described embodiment of the invention
Limitation, other any Spirit Essences without departing from the present invention and the change made under principle, modification, replacement, combine, simplification,
Equivalent substitute mode is should be, is included within protection scope of the present invention.
Claims (7)
1. a kind of traffic route sign based on saliency and deep learning, it is characterised in that including as follows
Step：
S1, download training dataset；
S2, pretreatment, are pre-processed to the image in data set in step S1；
S3, training convolutional neural networks model；
S4, conspicuousness are extracted；Region contrast method is employed, the road signs in test pictures are extracted；
S5, the road signs picture extracted in step S4 is input to the convolutional neural networks mould trained in step S3
In type, the classification of traffic route sign image is gone out by convolutional neural networks Model Identification.
2. a kind of traffic route sign based on saliency and deep learning according to claim 1,
Characterized in that, training dataset is downloaded in step S1, data set national fuzzy image processing contest and content from 2015
The Traffic Sign Recognition training set in contest is analyzed, the traffic mark being related to is divided into Warning Sign, Prohibition Sign, indicateing arm
Knowledge, way-finding sign, tourist district's mark, 7 major classes of road construction secure ID and accessory ID；The quality identified in image includes
Clearly, obscure, block, shade and tilt 5 types.
3. a kind of traffic route sign based on saliency and deep learning according to claim 1,
Characterized in that, step S2 pretreatments detailed process is as follows：
Each image, scales the images to 256 × 256 sizes, and then interception 5 is small greatly at image upper, middle and lower or so 5 is
224 × 224 standard size image, and the classification of standard size image corresponding thereto is preserved in pairs, " a standard size
Image-classification " is to being used as a data.
4. a kind of traffic route sign based on saliency and deep learning according to claim 1,
Characterized in that, step S3 convolutional neural networks workflows are as follows, input layer is made up of 32 × 32 sensing nodes, receives former
Beginning image A, then, calculation process between convolution and sub-sample alternately, it is described in detail below：
First hidden layer carries out convolution, and it is made up of 8 Feature Mappings, and each Feature Mapping is made up of 28 × 28 neurons,
Each neuron specifies the receiving of one 5 × 5；
Second hidden layer realizes sub-sample and local average, and it is equally made up of 8 Feature Mappings, but each of which Feature Mapping by
14 × 14 neuron compositions；Each neuron has the acceptance region of one 2 × 2, and one can train biasing and a sigmoid
Activation primitive；The operating point of coefficient and biasing control neuron can be trained；
3rd hidden layer carries out second of convolution, and it is made up of 20 Feature Mappings, and each Feature Mapping is by 10 × 10 nerves
Member composition；Each neuron in the hidden layer may have the cynapse being connected with the several Feature Mappings of next hidden layer to connect
Connect, it to first similar mode of convolutional layer to operate；
4th hidden layer carries out second of sub-sample and local average computation；It is made up of 20 Feature Mappings, but each feature
Mapping be made up of 5 × 5 neurons, and it to similar mode of sampling for the first time to operate；
5th hidden layer realizes the final stage of convolution, and it is made up of 120 neurons, and each neuron specifies one 5 × 5
Acceptance region；
It is finally a full articulamentum, obtains output vector.
5. a kind of traffic route sign based on saliency and deep learning according to claim 1,
Characterized in that, step S4 carries out conspicuousness processing to test pictures, region contrast method, i.e. RC methods are employed, by sky
Between relation be incorporated into area grade contrast calculate among；It is as follows that conspicuousness extraction step is carried out with RC methods：
S4.1, original image A will be inputted first be divided into color contrast on several target areas, then zoning grade；
Input picture is divided into several regions using a kind of image partition method based on figure, is that color histogram is set up in each region
Figure, for a region rk, its conspicuousness is calculated by measuring it with the color contrast in other regions in image, such as
Under：
<mrow>
<mi>S</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>r</mi>
<mi>k</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munder>
<mo>&Sigma;</mo>
<msub>
<mi>r</mi>
<mrow>
<mi>k</mi>
<mo>&NotEqual;</mo>
<msub>
<mi>r</mi>
<mi>i</mi>
</msub>
</mrow>
</msub>
</munder>
<mi>w</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>r</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<msub>
<mi>D</mi>
<mi>r</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>r</mi>
<mi>k</mi>
</msub>
<mo>,</mo>
<msub>
<mi>r</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
Wherein w (ri) it is region riWeights, Dr(rk,ri) it is two interregional color measurements；riIn number of pixels w (ri)
To emphasize the color contrast with bigger region；Region r1And r2Between color distance be defined as foloows：
<mrow>
<msub>
<mi>D</mi>
<mi>r</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>r</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<msub>
<mi>r</mi>
<mn>2</mn>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<msub>
<mi>n</mi>
<mn>1</mn>
</msub>
</munderover>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<msub>
<mi>n</mi>
<mn>2</mn>
</msub>
</munderover>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>c</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>c</mi>
<mn>2</mn>
</msub>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
<mi>D</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>c</mi>
<mn>1</mn>
</msub>
<mo>,</mo>
<mi>i</mi>
<mo>,</mo>
<msub>
<mi>c</mi>
<mn>2</mn>
</msub>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</mrow>
In order to preferably reflect a difference between color and primary color, we appear in this region using this color
In frequency be used as the weights of this color, wherein f (ck, i) it is i-th of color ck,iIn k-th of region rkIn all nk
The frequency of occurrences in individual color；
S4.2, the region contrast for being re-introduced on the basis of step S4.1 spatial weighting；Introduce one and include spatial information
Weight, the setting of weights is defined according to regional space distance, and the target area of space length farther out is endowed less
Weights, the nearer target area of space length is endowed larger weights；Increased with this immediate area influence and reduce compared with
The influence of far region；For any region rk, the conspicuousness based on spatial weighting region contrast is defined as follows：
<mrow>
<mi>S</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>r</mi>
<mi>k</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munder>
<mi>&Sigma;</mi>
<msub>
<mi>r</mi>
<mrow>
<mi>k</mi>
<mo>&NotEqual;</mo>
<msub>
<mi>r</mi>
<mi>i</mi>
</msub>
</mrow>
</msub>
</munder>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<msub>
<mi>D</mi>
<mi>s</mi>
</msub>
<mo>(</mo>
<mrow>
<msub>
<mi>r</mi>
<mrow>
<mi>k</mi>
<mo>,</mo>
</mrow>
</msub>
<msub>
<mi>r</mi>
<mi>i</mi>
</msub>
</mrow>
<mo>)</mo>
<mo>/</mo>
<msubsup>
<mi>&sigma;</mi>
<mi>s</mi>
<mn>2</mn>
</msubsup>
<mo>)</mo>
</mrow>
<mi>w</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>r</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<msub>
<mi>D</mi>
<mi>r</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>r</mi>
<mi>k</mi>
</msub>
<mo>,</mo>
<msub>
<mi>r</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
Wherein, Ds(rk,ri) it is region rkAnd riBetween space length, σsControl the intensity of space weights；Larger σsValue more can
Reduce the influence of space weights so that region farther out is more conducive to obtain gray-scale map after the conspicuousness of current region, processing
As B；
S4.3, the salient region for extracting test image, obtain including the gray level image B of traffic route identified areas, connect down
Binary image C is obtained to carry out binaryzation to gray level image B；
Holes filling in S4.4, the binary image C for obtaining step S4.3, obtains image D, then using image D as mask,
Retain the part that D correspondence positions are 1 in original image A, other are set to 0, obtain including the aobvious of traffic route mark region
Work property image.
6. a kind of traffic route sign based on saliency and deep learning according to claim 5,
Characterized in that, binary image uses global threshold algorithm in step S4.3, Otsu methods carry out binary conversion treatment to image；
Otsu is the statistical property based on entire image, realizes the automatic selection of threshold value；If image has M gray value, span is 0
~M-1, chooses gray value t, divides the image into two groups of G within this range0And G1, G0Comprising pixel gray value in 0~t, G1
Gray value in t+1~M-1, total number of image pixels, n are represented with NiNumber of the gray value for i pixel is represented, then each ash
The Probability p that angle value i occursiFor：
<mrow>
<msub>
<mi>p</mi>
<mi>i</mi>
</msub>
<mo>=</mo>
<mfrac>
<msub>
<mi>n</mi>
<mi>i</mi>
</msub>
<mi>N</mi>
</mfrac>
<mo>;</mo>
</mrow>
G0Class occur probability and average be：
<mrow>
<msub>
<mi>w</mi>
<mn>0</mn>
</msub>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mi>t</mi>
</munderover>
<msub>
<mi>p</mi>
<mi>i</mi>
</msub>
</mrow>
<mrow>
<msub>
<mi>u</mi>
<mn>0</mn>
</msub>
<mo>=</mo>
<msubsup>
<mi>&Sigma;</mi>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mi>t</mi>
</msubsup>
<msub>
<mi>ip</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
</mrow>
2
Wherein w0For G0The probability that class occurs, u0For G0The average of class；
G1Class occur probability and average be：
<mrow>
<msub>
<mi>w</mi>
<mn>1</mn>
</msub>
<mo>=</mo>
<munderover>
<mi>&Sigma;</mi>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mi>i</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
<mrow>
<mi>M</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</munderover>
<msub>
<mi>p</mi>
<mi>i</mi>
</msub>
<mo>=</mo>
<mn>1</mn>
<mo>-</mo>
<msub>
<mi>w</mi>
<mn>0</mn>
</msub>
</mrow>
<mrow>
<msub>
<mi>u</mi>
<mn>1</mn>
</msub>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mi>i</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
<mrow>
<mi>M</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</munderover>
<msub>
<mi>ip</mi>
<mi>i</mi>
</msub>
</mrow>
Wherein w1For G1The probability that class occurs, u1For G1The average of class；
Then inter-class variance δ (t)2=w0w1(u0-u1)2
Global threshold T is the value for the t for making inter-class variance maximum, is：
T=argmax δ (t), t ∈ [0, M-1]
Entire image is divided into foreground and background two parts by threshold value T, the gray value of every pixel is more than being set as this threshold value
225, it is set as 0 less than this threshold value；Gray level image is processed into only black and white binary image, two-value is finally given
Change image C.
7. a kind of traffic route sign based on saliency and deep learning according to claim 1,
Characterized in that, step S5 is according to training model measurement image recognition effect：Included after conspicuousness in step S4 is extracted
Test experiments are carried out in the convolutional neural networks trained in the picture input step S3 for having road signs region.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710122736.XA CN107122701A (en) | 2017-03-03 | 2017-03-03 | A kind of traffic route sign based on saliency and deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710122736.XA CN107122701A (en) | 2017-03-03 | 2017-03-03 | A kind of traffic route sign based on saliency and deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107122701A true CN107122701A (en) | 2017-09-01 |
Family
ID=59718198
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710122736.XA Pending CN107122701A (en) | 2017-03-03 | 2017-03-03 | A kind of traffic route sign based on saliency and deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107122701A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107703936A (en) * | 2017-09-22 | 2018-02-16 | 南京轻力舟智能科技有限公司 | Automatic Guided Vehicle system and dolly localization method based on convolutional neural networks |
| CN108154102A (en) * | 2017-12-21 | 2018-06-12 | 安徽师范大学 | A kind of traffic sign recognition method |
| CN108256467A (en) * | 2018-01-15 | 2018-07-06 | 河北科技大学 | A kind of method for traffic sign detection of view-based access control model attention mechanism and geometric properties |
| CN109102493A (en) * | 2018-07-03 | 2018-12-28 | 柳州市木子科技有限公司 | A kind of automobile metal plate work flaw detection system based on CNN and LR |
| CN109636881A (en) * | 2018-12-19 | 2019-04-16 | 沈阳天择智能交通工程有限公司 | Based on AI identification technology traffic accident situ sketch drafting method |
| CN109711399A (en) * | 2018-11-05 | 2019-05-03 | 北京三快在线科技有限公司 | Shop recognition methods based on image, device, electronic equipment |
| CN110060238A (en) * | 2019-04-01 | 2019-07-26 | 桂林电子科技大学 | Pcb board based on deep learning marks print quality inspection method |
| CN110321905A (en) * | 2019-07-11 | 2019-10-11 | 广东工业大学 | Abnormal area detection method, system and associated component based on semantic segmentation |
| CN111257326A (en) * | 2020-01-22 | 2020-06-09 | 重庆大学 | Metal processing area extraction method |
| CN112734733A (en) * | 2021-01-12 | 2021-04-30 | 天津大学 | Non-reference image quality monitoring method based on channel recombination and feature fusion |
| WO2022078216A1 (en) * | 2020-10-14 | 2022-04-21 | 华为云计算技术有限公司 | Target recognition method and device |
| CN114973207A (en) * | 2022-08-01 | 2022-08-30 | 成都航空职业技术学院 | Road sign identification method based on target detection |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103955718A (en) * | 2014-05-15 | 2014-07-30 | 厦门美图之家科技有限公司 | Image subject recognition method |
| CN105590319A (en) * | 2015-12-18 | 2016-05-18 | 华南理工大学 | Method for detecting image saliency region for deep learning |
| CN105931255A (en) * | 2016-05-18 | 2016-09-07 | 天津工业大学 | Method for locating target in image based on obviousness and deep convolutional neural network |
-
2017
- 2017-03-03 CN CN201710122736.XA patent/CN107122701A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103955718A (en) * | 2014-05-15 | 2014-07-30 | 厦门美图之家科技有限公司 | Image subject recognition method |
| CN105590319A (en) * | 2015-12-18 | 2016-05-18 | 华南理工大学 | Method for detecting image saliency region for deep learning |
| CN105931255A (en) * | 2016-05-18 | 2016-09-07 | 天津工业大学 | Method for locating target in image based on obviousness and deep convolutional neural network |
Non-Patent Citations (4)
| Title |
|---|
| L.M. RASDI RERE 等: "Metaheuristic Algorithms for Convolution Neural Networks", 《ARXIV》 * |
| 景辉芳: "深度神经网络的研究及其在植物叶片图像识别中的应用", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 朱驰 等: "改进图割的显著性区域检测算法", 《计算机工程与设计》 * |
| 黄琳: "基于深层神经网络的交通标志识别方法研究", 《万方数据知识服务平台》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107703936A (en) * | 2017-09-22 | 2018-02-16 | 南京轻力舟智能科技有限公司 | Automatic Guided Vehicle system and dolly localization method based on convolutional neural networks |
| CN108154102A (en) * | 2017-12-21 | 2018-06-12 | 安徽师范大学 | A kind of traffic sign recognition method |
| CN108256467B (en) * | 2018-01-15 | 2020-06-05 | 河北科技大学 | Traffic sign detection method based on visual attention mechanism and geometric features |
| CN108256467A (en) * | 2018-01-15 | 2018-07-06 | 河北科技大学 | A kind of method for traffic sign detection of view-based access control model attention mechanism and geometric properties |
| CN109102493A (en) * | 2018-07-03 | 2018-12-28 | 柳州市木子科技有限公司 | A kind of automobile metal plate work flaw detection system based on CNN and LR |
| CN109711399A (en) * | 2018-11-05 | 2019-05-03 | 北京三快在线科技有限公司 | Shop recognition methods based on image, device, electronic equipment |
| CN109711399B (en) * | 2018-11-05 | 2021-04-27 | 北京三快在线科技有限公司 | Shop identification method and device based on image and electronic equipment |
| CN109636881A (en) * | 2018-12-19 | 2019-04-16 | 沈阳天择智能交通工程有限公司 | Based on AI identification technology traffic accident situ sketch drafting method |
| CN110060238A (en) * | 2019-04-01 | 2019-07-26 | 桂林电子科技大学 | Pcb board based on deep learning marks print quality inspection method |
| CN110060238B (en) * | 2019-04-01 | 2022-04-19 | 桂林电子科技大学 | PCB label printing quality detection method based on deep learning |
| CN110321905A (en) * | 2019-07-11 | 2019-10-11 | 广东工业大学 | Abnormal area detection method, system and associated component based on semantic segmentation |
| CN111257326A (en) * | 2020-01-22 | 2020-06-09 | 重庆大学 | Metal processing area extraction method |
| CN111257326B (en) * | 2020-01-22 | 2021-02-26 | 重庆大学 | Metal processing area extraction method |
| WO2022078216A1 (en) * | 2020-10-14 | 2022-04-21 | 华为云计算技术有限公司 | Target recognition method and device |
| CN112734733A (en) * | 2021-01-12 | 2021-04-30 | 天津大学 | Non-reference image quality monitoring method based on channel recombination and feature fusion |
| CN112734733B (en) * | 2021-01-12 | 2022-11-01 | 天津大学 | Non-reference image quality monitoring method based on channel recombination and feature fusion |
| CN114973207A (en) * | 2022-08-01 | 2022-08-30 | 成都航空职业技术学院 | Road sign identification method based on target detection |
| CN114973207B (en) * | 2022-08-01 | 2022-10-21 | 成都航空职业技术学院 | Road sign identification method based on target detection |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107122701A (en) | A kind of traffic route sign based on saliency and deep learning | |
| CN110163187B (en) | F-RCNN-based remote traffic sign detection and identification method | |
| CN109977812B (en) | Vehicle-mounted video target detection method based on deep learning | |
| CN109165623B (en) | Rice disease spot detection method and system based on deep learning | |
| CN109800736A (en) | A kind of method for extracting roads based on remote sensing image and deep learning | |
| CN103034863B (en) | The remote sensing image road acquisition methods of a kind of syncaryon Fisher and multiple dimensioned extraction | |
| CN110619282A (en) | Automatic extraction method for unmanned aerial vehicle orthoscopic image building | |
| CN109949316A (en) | A kind of Weakly supervised example dividing method of grid equipment image based on RGB-T fusion | |
| CN106326858A (en) | Road traffic sign automatic identification and management system based on deep learning | |
| CN107609602A (en) | A kind of Driving Scene sorting technique based on convolutional neural networks | |
| CN101814144B (en) | Water-free bridge target identification method in remote sensing image | |
| CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
| CN105975931A (en) | Convolutional neural network face recognition method based on multi-scale pooling | |
| CN108491797A (en) | A kind of vehicle image precise search method based on big data | |
| CN109598241A (en) | Satellite image marine vessel recognition methods based on Faster R-CNN | |
| CN109214308A (en) | A kind of traffic abnormity image identification method based on focal loss function | |
| CN109325438A (en) | The real-time identification method of live panorama traffic sign | |
| CN107169985A (en) | A kind of moving target detecting method based on symmetrical inter-frame difference and context update | |
| CN108154102A (en) | A kind of traffic sign recognition method | |
| CN109543632A (en) | A kind of deep layer network pedestrian detection method based on the guidance of shallow-layer Fusion Features | |
| CN109766936A (en) | Image change detection method based on information transmitting and attention mechanism | |
| CN108629368B (en) | Multi-modal foundation cloud classification method based on joint depth fusion | |
| CN104504389A (en) | Satellite cloud amount computing method based on convolution neural network | |
| CN103984963B (en) | Method for classifying high-resolution remote sensing image scenes | |
| CN107038416A (en) | A kind of pedestrian detection method based on bianry image modified HOG features |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170901 |
|
| RJ01 | Rejection of invention patent application after publication |