CN105354568A - Convolutional neural network based vehicle logo identification method - Google Patents

Abstract

The invention provides a convolutional neural network based vehicle logo identification method. The method is specifically implemented by the following steps of: (1) inputting a to-be-detected picture shot by a high-resolution camera device in a traffic intersection; (2) positioning a vehicle logo; (3) constructing and training a convolutional neural network; and (4) identifying the vehicle logo. With the adoption of the convolutional neural network (CNN) based vehicle logo identification method, the shortcomings of complicated extraction feature operator, poor timeliness and complicated model in the prior art can be effectively overcome, and the calculation amount is effectively reduced; and features of CNN self-learning have higher robustness on environmental change, so that the identification rate of the vehicle logo is increased.

Description

Based on the automobile logo identification method of convolutional neural networks

Technical field

The invention belongs to technical field of image processing, further relate to a kind of automobile logo identification method based on convolutional neural networks of the mode identification technology based on image.The present invention is directed in traffic system, the vehicle pictures of the high definition photographing device gained arranged by crossing, carries out the vehicle-logo location of automobile, and then identifies the car mark of automobile, realizes locating the robotization of car target and identifying.

Background technology

Along with improving constantly of social economy's level and popularizing of vehicle, the demand of the communication that scale constantly expands to more intelligentized technology and system is larger, and intelligent transportation system has become the hot issue of social life.Vehicle identification system is as the important component part of intelligent transportation system, and in expressway access, parking lot unattended, the vehicles peccancy field such as automatically to record has this to apply widely, and its realization has very large economic worth and realistic meaning.

Vehicle-logo recognition is an importance of vehicle identification.Vehicle-logo recognition technology refers to digital picture or video signal flow for object, by image procossing and automatic identifying method, obtains a kind of practical technique of motor vehicles brand message.Vehicle-logo recognition system comprises car target location and vehicle-logo recognition binomial gordian technique.The features such as the otherness under the diversity had due to car sample body and varying environment condition, add car target locational uncertainty in the pictorial information that artificial shooting obtains, therefore find a kind of outstanding vehicle-logo location and recognition methods multi-crossed disciplines and challenging technical matters.

The method of existing vehicle-logo location, mostly adopt the method for rim detection and grey level histogram template matches, because car mark is little, these class methods are easily subject to the impact of background environment.The method of some vehicle-logo recognition is suggested, particularly use the more recognition methods based on histograms of oriented gradients HOG characteristic sum support vector machines sorter at present, major part is all based on car plate and car target relative position determination car cursor position, then extract car target histograms of oriented gradients HOG feature, utilize support vector machines to be trained to sorter and carry out vehicle-logo recognition.In vehicle-logo recognition, histograms of oriented gradients HOG adds support vector machines algorithm owing to have employed histograms of oriented gradients HOG feature, and histograms of oriented gradients HOG descriptor generative process is tediously long, cause speed slow, poor real, due to the character of gradient, this descriptor is quite responsive to noise.Existing most of vehicle-logo recognition algorithm, process is complicated, and calculated amount is too large, and discrimination is not high, is easily subject to the impact of environmental baseline, so need the proposition of new research method.

In recent years, along with the development of large data, degree of depth Learning Studies, convolutional neural networks CNN has become the study hotspot of current speech analysis and field of image recognition, its weights shared network structure makes it more to be similar to biological neural network, reduce the complexity of network model, decrease the quantity of weights.It is more obvious that this advantage shows when the input of network is multidimensional image, makes image directly as the input of network, can avoid feature extraction complicated in tional identification algorithm and data reconstruction processes.Convolutional network is a multilayer perceptron for identifying two-dimensional shapes and particular design, and the distortion of this network structure to translation, proportional zoom, inclination or his form altogether has height unchangeability.

D.F.Llorca, R.Arroyo, the method of a set of vehicle-logo recognition based on histograms of oriented gradients HOG and support vector machines is proposed in paper " VehiclelogorecognitionintrafficimagesusingHOGfeaturesand SVM " (Proceedingsofthe16thInternationalIEEEAnnualConferenceonI ntelligentTransportationSystems, 2013) that M.A.Sotelo delivers at it.The method is License Plate first, car mark is utilized to be in priori directly over car plate, above car plate, use moving window to shift to an earlier date candidate target region, then extract candidate region histograms of oriented gradients HOG feature, the sorter finally utilizing support vector machines to train carries out the classification of car mark.The weak point that the method exists is, one, and owing to the process employs histograms of oriented gradients HOG feature, histograms of oriented gradients HOG descriptor generative process is tediously long, causes speed slow, poor real.Its two, due to the character of the method gradient, histograms of oriented gradients HOG descriptor is quite responsive to noise, is easily subject to the interference of noise.

Propose a kind of car mark based on pattern-recognition in the patent " a kind of car mark based on pattern-recognition is located and recognition methods automatically " (number of patent application: CN201410367377, publication number: CN104182728A) of Pci-Suntek Technology Co., Ltd.'s application automatically to locate and recognition methods.First the method utilizes car plate detection technique, obtain size and the position of car plate, thus according to car plate and car target relative position, carry out car target just to locate, secondly the strong classifier Adaboost algorithm based on Ha Er Haar feature is utilized to carry out car target secondary location, obtain some doubtful car target areas, the support vector machines algorithm based on histograms of oriented gradients HOG feature is again utilized to screen doubtful car mark region, choose there is maximum confidence region as vehicle-logo location result, the support vector machines algorithm based on HOG feature is finally utilized to carry out the identification of car target.The weak point that the method exists is, have employed based on the strong classifier Adaboost algorithm of Ha Er Haar feature and the support vector machines algorithm based on histograms of oriented gradients HOG feature in positioning flow, the support vector machines algorithm based on histograms of oriented gradients HOG feature is have employed in vehicle-logo recognition flow process, altogether have employed three sorters, considerably increase computation complexity.And HOG descriptor generative process length consuming time, cause speed slow, poor real.

Patent " car mark automatic identifying method and the system " (number of patent application: CN201310170528 of Shanghai Communications University's application, publication number: CN103279738A) a kind of car mark automatic identifying method of middle proposition, comprise off-line training subsystem and ONLINE RECOGNITION subsystem.Intensive Scale invariant features transform dense-SIFT, according to the correlativity of intensive Scale invariant features transform dense-SIFT and visual word, is mapped to all visual word and represents by the method, increases feature interpretation.Adopt support vector machine training cart mark sorter, realize vehicle-logo recognition.The weak point that the method exists is that, owing to have employed intensive Scale invariant features transform dense-SIFT feature operator, dimension is high, and computing time is long, poor real.

Summary of the invention

The object of the invention is the deficiency existed for above-mentioned prior art, propose a kind of automobile logo identification method based on convolutional neural networks.The present invention's calculated amount compared with other vehicle-logo recognition technology in prior art is little, and accuracy is high, strong adaptability.

The concrete steps that the present invention realizes comprise as follows:

1., based on an automobile logo identification method for convolutional neural networks, comprise the steps:

(1) the car mark picture to be detected that in traffic intersection, high definition photographing device obtains is inputted;

(2) vehicle-logo location:

(2a) binaryzation operation is carried out to the car mark picture to be detected of input, obtain the car after binaryzation and to mark on a map sheet;

(2b) car after binaryzation is marked on a map the morphological operation that sheet corrodes and expand, obtain UNICOM region;

(2c) in UNICOM region, utilize the ratio of width to height and the rectangular characteristic in car plate UNICOM region, filter out car plate UNICOM region, obtain top-left coordinates (x1, y1) and the lower right coordinate (x2, y2) in car plate UNICOM region;

(2d) mark on a map in sheet at measuring car to be checked, intercept car mark areal map by sliding window, sliding window base is initial with car plate upper edge, and up sliding and intercept candidate region 3 times in the Central Line along car plate, obtains car mark areal map;

(3) also training convolutional neural networks CNN is built:

(3a) build the convolutional neural networks CNN containing 7 layers, 7 layers is convolutional layer Conv1 successively, pond layer Pool2, convolutional layer Conv3, pond layer Pool4, full articulamentum Fc5, full articulamentum Fc6, classification layer Softmax7;

(3b) input has marked and the car mark area sample picture of gray processing, and training convolutional neural networks CNN, until the loss function J (θ)≤0.0001 of output layer, obtains the convolutional neural networks CNN of vehicle-logo recognition;

(4) vehicle-logo recognition:

(4a) gray processing operation is carried out to car mark areal map;

(4b) the car mark areal map resolution after gray processing is normalized to 38 × 38 pixel sizes, obtains the car after processing and mark on a map;

(4c) by process after car mark on a map input vehicle-logo recognition convolutional neural networks CNN, final Output rusults.

The present invention compared with prior art has the following advantages:

The first, because the present invention adopts first positioning licence plate region, then the vehicle-logo location method of region with sliding window intercepting car mark region is just being gone up at car plate, overcome extracting directly car mark in prior art and be subject to background environment impact, can not, by the problem accurately extracted, make the present invention can extract car mark region exactly from complex background environment.

The second, because the present invention adopts the automobile logo identification method based on convolutional neural networks CNN, by the network self study feature of multilayer in convolutional neural networks CNN, avoid the process needing engineer's feature in tional identification algorithm, and the feature of convolutional neural networks CNN self study has higher robustness to environmental change, make to present invention reduces calculated amount and complicacy, improve discrimination, to complex background, there is stronger adaptability.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is the schematic diagram of vehicle-logo location of the present invention;

Fig. 3 is convolutional neural networks CNN structural drawing of the present invention;

Fig. 4 is that the car of portion markings of the present invention is marked on a map.

Embodiment

Below in conjunction with accompanying drawing, the present invention will be further described.

With reference to Fig. 1, the concrete steps that the present invention realizes are as follows:

Step 1, the picture to be detected that in input traffic intersection, high definition photographing device takes.

Picture to be detected comprises apparent mark car plate and car mark, and standard car plate area size is 180 × 60 pixels.

Step 2, vehicle-logo location.

Carry out binaryzation operation to the picture to be detected of input, obtain the picture after binaryzation, concrete operations are:

The first step, chooses the turquoise rgb value of primary colors red of car plate background color in 50 marker samples, the average of the turquoise rgb value of statistics primary colors red;

Second step, according to the following formula, carries out binaryzation operation to the car mark picture to be detected of input:

Wherein, q _ijrepresent the gray-scale value of the pixel of the picture after binaryzation, r ₀, g ₀, b ₀represent the turquoise average of primary colors red of samples pictures pixel respectively, r _ij,g _ij, b _ijrepresent the turquoise value of primary colors red of picture pixels to be detected point respectively, i, j represent line number and the columns of picture pixels point respectively.

To the morphological operation that the picture after binaryzation corrodes and expands, obtain UNICOM region.In UNICOM region, utilize the ratio of width to height in car plate UNICOM region and the ratio of width to height of rectangular characteristic and car plate to be 3:1, car plate is wider than height, and be less than 5 degree with the horizontal pitch angle of shooting road, filter out car plate UNICOM region as Fig. 2 (a), obtain the top-left coordinates (x in car plate UNICOM region _1,y ₁) and lower right coordinate (x _2,y ₂).

In picture to be detected, intercept car mark areal map with sliding window, the window of sliding window is square, and its length of side is step-length is wherein, l represents the length of side of sliding window window, and h represents the step-length of sliding window window, (x _1,y ₁) represent the top-left coordinates in car plate UNICOM region, (x _2,y ₂) represent the lower right coordinate in car plate UNICOM region.Sliding window is initial with car plate upper edge below, and up slide and intercept candidate region 3 times in the Central Line along car plate, as shown in Fig. 2 (b), the final car mark areal map exported is as Fig. 2 (c).

Step 3, builds and training convolutional neural networks.

Build the convolutional neural networks CNN containing 7 layers as shown in Figure 3,7 layers is convolutional layer Conv1 successively, pond layer Pool2, convolutional layer Conv3, pond layer Pool4, full articulamentum Fc5, full articulamentum Fc6, classification layer Softmax7, the convolutional neural networks CNN concrete steps built containing the vehicle-logo recognition of 7 layers are as follows:

1st step, by the car mark areal map of 38 × 38 pixel sizes input convolutional layer Conv1, to it, to carry out block size be 5 × 5 pixels and step-length is the convolution operation of 1 pixel, altogether uses 32 convolution kernels, obtains the characteristic pattern of 32 34 × 34 pixel sizes;

32 characteristic patterns that convolutional layer Conv1 exports are input to pond layer Pool2 by the 2nd step, and carry out the operation of maximum pondization to it, the size of pond block is 2 × 2 pixels, and step-length is 1 pixel, obtains the characteristic pattern that 32 resolution are 17 × 17 pixels;

3rd step, 32 characteristic patterns input convolutional layer Conv3 that pond layer Pool2 is exported, to it, to carry out block size be 5 × 5 pixels and step-length is the convolution operation of 1 pixel, altogether uses 64 convolution kernels, obtains the characteristic pattern that 64 resolution are 13 × 13 pixels;

4th step, 64 characteristic pattern input pond layer Pool4 that convolutional layer Conv3 is exported, carry out the operation of maximum pondization to it, the size of pond block is 2 × 2 pixels, and step-length is 1 pixel, obtains the characteristic pattern that 64 resolution are 7 × 7 pixels;

5th step, 64 characteristic patterns that pond layer Pool4 exports are inputted full articulamentum Fc5, according to the following formula, wherein each pixel is activated, obtain the value of the pixel of the characteristic pattern after activating, characteristic pattern after activating is arranged in 1 dimensional vector with the order of row, obtains the proper vector of 1 × 3136 dimension:

$f\left(x\right)=\frac{e^x-e^{-x}}{e^x+e^{-x}}$

Wherein, f (x) represents the value of the pixel of the characteristic pattern after activating, and x represents the value of the pixel activating front characteristic pattern, and e represents a natural constant infinitely do not circulated, and value is 2.7182;

6th step, inputs full articulamentum Fc6 by the proper vector that full articulamentum Fc5 exports, forms general neural network, and output is the proper vector of 1 × 500 dimension;

7th step, the proper vector input classification layer Softmax7 that full articulamentum Fc6 is exported, obtain the tag along sort of car mark areal map, this layer of meeting calculates the probability of often kind of tag along sort, and the label of maximum probability is exported, wherein the expectation function of softmax classification is as follows:

$h_{\mathrm{\θ}}\left({\mathrm{\α}}^{\left(i\right)}\right)=\left[\begin{array}{c}p\left({\mathrm{\β}}^{\left(i\right)}=1|{\mathrm{\α}}^{\left(i\right)};\mathrm{\θ}\right)\\ p\left({\mathrm{\β}}^{\left(i\right)}=2|{\mathrm{\α}}^{\left(i\right)};\mathrm{\θ}\right)\\ .\\ .\\ .\\ p\left({\mathrm{\β}}^{\left(i\right)}=k|{\mathrm{\α}}^{\left(i\right)};\mathrm{\θ}\right)\end{array}\right]$

Wherein, h _θ(α) expectation function that softmax classifies is represented, α represents the proper vector that in convolutional neural networks CNN, full articulamentum Fc6 exports, β represents the label corresponding with the proper vector α that articulamentum Fc6 complete in convolutional neural networks CNN exports, p (β=t| α; When θ) representing the proper vector α being input as full articulamentum Fc6 output in convolutional neural networks CNN, label β equals the probability of t, t ∈ 1,2 ..., k, θ represent model parameter and θ ₁, θ ₂..., θ _k∈ R ⁿ⁺¹, softmax Classification Loss function is as follows:

$J\left(\mathrm{\θ}\right)=-\frac{1}{m}\[\underset{i=1}{\overset{m}{\Σ}}{\mathrm{\β}}^{\left(i\right)}{\mathrm{logh}}_{\mathrm{\θ}}\left({\mathrm{\α}}^{\left(i\right)}\right)+(1-{\mathrm{\β}}^{\left(i\right)})log(1-h_{\mathrm{\θ}}({\mathrm{\α}}^{\left(i\right)}\left)\right)\]$

Wherein, J (θ) represents loss function, and m represents the quantity of car mark areal map sample, h _θ(α) expectation function that softmax classifies is represented, α represents the proper vector that in convolutional neural networks CNN, full articulamentum Fc6 exports, β represents the label corresponding with the proper vector α that articulamentum Fc6 complete in convolutional neural networks CNN exports, and θ represents model parameter.

Input has marked and the car mark area sample picture of gray processing, training convolutional neural networks CNN, and the process of training is as follows:

The first step, in the forward direction stage, gets a sample from sample set, and sample is inputted convolutional neural networks CNN and calculate corresponding actual output, in this stage, information through conversion step by step, is sent to output layer from input layer;

Second step, in the back-propagation stage, calculates the actual difference exporting the ideal corresponding with sample label and export, by the method for minimization error, and backpropagation adjustment weight matrix;

3rd step, repeats the operation of the first step and second step, and the loss function J (θ)≤0.0001 of the output until convolutional neural networks CNN classifies after layer Softmax7, obtains the convolutional neural networks CNN of vehicle-logo recognition.

Step 4, vehicle-logo recognition.

To car mark areal map, carry out gray processing operation, and car mark areal map resolution is contracted to 38 × 38 pixel sizes, obtain the car after processing and mark on a map, as shown in Figure 4.Car after process is marked on a map and inputs the convolutional neural networks CNN of vehicle-logo recognition, final Output rusults.

Below in conjunction with emulation experiment, effect of the present invention is further described.

1, emulation experiment condition:

The present invention's database used be collect and make one group comprise 10 class car target vehicle mark bases, wherein each car indicates 3600 training plans and 240 test patterns, totally 36000 training plans and 2400 test patterns, negative sample is 120,000, comprises various non-car target picture.Hardware platform is: IntelCore2DuoCPUE65502.33GHZ, 3GBRAM, software platform: vs2010, MATLABR2012a.

2, experiment content and result:

The ten kind car marks of the present invention first in collecting cart mark database, totally 36000 training datas and 2400 test datas.By License Plate, car mark is just located accurately to locate with car mark and is finally accurately determined car target area, is input in the convolutional neural networks CNN sorter trained by train mark areal map and classifies, obtain a result.The simulation result of contrast following table, in 2400 data of test, each classification Audi, Honda, BYD, mark, Buick, popular, Toyota, Jeep, Kia, Chang'an, number of errors is 6 respectively, 10,12,5,6,8,11,7,8,9.Can find out that the present invention has higher discrimination at identification car timestamp.

Car mark class Audi Honda BYD Mark Buick Error number 6 10 12 5 6 Car mark class Popular Toyota Jeep Kia Chang'an Error number 8 11 7 8 9

Car mark class Audi Honda BYD Mark Buick Discrimination 97.5％ 95.83％ 95％ 97.9％ 97.5％ Car mark class Popular Toyota Jeep Kia Chang'an Discrimination 96.67％ 95.42％ 97.18％ 96.67％ 96.25％

Claims (7)

1., based on an automobile logo identification method for convolutional neural networks, comprise the steps:

(1) the car mark picture to be detected that in traffic intersection, high definition photographing device obtains is inputted;

(2) vehicle-logo location:

(2a) binaryzation operation is carried out to the car mark picture to be detected of input, obtain the car after binaryzation and to mark on a map sheet;

(2b) car after binaryzation is marked on a map the morphological operation that sheet corrodes and expand, obtain UNICOM region;

(2c) in UNICOM region, utilize the ratio of width to height and the rectangular characteristic in car plate UNICOM region, filter out car plate UNICOM region, obtain the top-left coordinates (x in car plate UNICOM region _1,y ₁) and lower right coordinate (x _2,y ₂);

(2d) mark on a map in sheet at measuring car to be checked, intercept car mark areal map by sliding window, sliding window base is initial with car plate upper edge, and up sliding and intercept candidate region 3 times in the Central Line along car plate, obtains car mark areal map;

(3) also training convolutional neural networks CNN is built:

(3a) build the convolutional neural networks CNN containing 7 layers, 7 layers is convolutional layer Conv1 successively, pond layer Pool2, convolutional layer Conv3, pond layer Pool4, full articulamentum Fc5, full articulamentum Fc6, classification layer Softmax7;

(3b) input has marked and the car mark area sample picture of gray processing, and training convolutional neural networks CNN, until the loss function J (θ)≤0.0001 of output layer, obtains the convolutional neural networks CNN of vehicle-logo recognition;

(4) vehicle-logo recognition:

(4a) gray processing operation is carried out to car mark areal map;

(4b) the car mark areal map resolution after gray processing is normalized to 38 × 38 pixel sizes, obtains the car after processing and mark on a map;

(4c) by process after car mark on a map input vehicle-logo recognition convolutional neural networks CNN, final Output rusults.

2. the automobile logo identification method based on convolutional neural networks according to claim 1, it is characterized in that: the car mark picture to be detected described in step (1) comprises apparent mark car plate and car mark, and standard car plate area size is 180 × 60 pixels.

3. the automobile logo identification method based on convolutional neural networks according to claim 1, is characterized in that: described in step (2a) to carry out the step of binaryzation operation to the car mark picture to be detected of input as follows:

The first step, chooses the turquoise rgb value of primary colors red of car plate background color in 50 marker samples, the average of the turquoise rgb value of statistics primary colors red;

Second step, according to the following formula, carries out binaryzation operation to the car mark picture to be detected of input:

Wherein, q _ijrepresent the gray-scale value of the pixel of the picture after binaryzation, r ₀, g ₀, b ₀represent the turquoise average of primary colors red of samples pictures pixel respectively, r _ij,g _ij, b _ijrepresent the turquoise value of primary colors red of picture pixels to be detected point respectively, i, j represent line number and the columns of picture pixels point respectively.

4. the automobile logo identification method based on convolutional neural networks according to claim 1, is characterized in that:

The ratio of width to height and the rectangular characteristic in the car plate UNICOM region described in step (2c) refer to, the ratio of width to height of car plate is 3:1, and car plate is wider than height, and are less than 5 degree with the horizontal pitch angle of shooting road.

5. the automobile logo identification method based on convolutional neural networks according to claim 1, is characterized in that: the window of the sliding window described in step (2d) is square, and its length of side is: step-length is: wherein, l represents the length of side of sliding window window, and h represents the step-length of sliding window window, (x _1,y ₁) represent the top-left coordinates in car plate UNICOM region, (x _2,y ₂) represent the lower right coordinate in car plate UNICOM region.

6. the automobile logo identification method based on convolutional neural networks according to claim 1, is characterized in that: the concrete steps that step (3a) described structure contains the convolutional neural networks CNN of 7 layers are as follows:

1st step, by the car mark areal map of 38 × 38 pixel sizes input convolutional layer Conv1, to it, to carry out block size be 5 × 5 pixels and step-length is the convolution operation of 1 pixel, altogether uses 32 convolution kernels, obtains the characteristic pattern of 32 34 × 34 pixel sizes;

32 characteristic patterns that convolutional layer Conv1 exports are input to pond layer Pool2 by the 2nd step, and carry out the operation of maximum pondization to it, the size of pond block is 2 × 2 pixels, and step-length is 1 pixel, obtains the characteristic pattern that 32 resolution are 17 × 17 pixels;

3rd step, 32 characteristic patterns input convolutional layer Conv3 that pond layer Pool2 is exported, to it, to carry out block size be 5 × 5 pixels and step-length is the convolution operation of 1 pixel, altogether uses 64 convolution kernels, obtains the characteristic pattern that 64 resolution are 13 × 13 pixels;

4th step, 64 characteristic pattern input pond layer Pool4 that convolutional layer Conv3 is exported, carry out the operation of maximum pondization to it, the size of pond block is 2 × 2 pixels, and step-length is 1 pixel, obtains the characteristic pattern that 64 resolution are 7 × 7 pixels;

5th step, 64 characteristic patterns that pond layer Pool4 exports are inputted full articulamentum Fc5, according to the following formula, wherein each pixel is activated, obtain the value of the pixel of the characteristic pattern after activating, characteristic pattern after activating is arranged in 1 dimensional vector with the order of row, obtains the proper vector of 1 × 3136 dimension:

$f\left(x\right)=\frac{e^x-e^{-x}}{e^x+e^{-x}}$

Wherein, f (x) represents the value of the pixel of the characteristic pattern after activating, and x represents the value of the pixel activating front characteristic pattern, and e represents a natural constant infinitely do not circulated, and value is 2.7182;

6th step, inputs full articulamentum Fc6 by the proper vector that full articulamentum Fc5 exports, forms general neural network, and output is the proper vector of 1 × 500 dimension;

7th step, the proper vector input classification layer Softmax7 that full articulamentum Fc6 is exported, obtain the tag along sort of car mark areal map, this layer of meeting calculates the probability of often kind of tag along sort, and the label of maximum probability is exported, wherein the expectation function of softmax classification is as follows:

$h_{\mathrm{\θ}}\left({\mathrm{\α}}^{\left(i\right)}\right)=\left[\begin{array}{c}p\left({\mathrm{\β}}^{\left(i\right)}=1|{\mathrm{\α}}^{\left(i\right)};\mathrm{\θ}\right)\\ p\left({\mathrm{\β}}^{\left(i\right)}=2|{\mathrm{\α}}^{\left(i\right)};\mathrm{\θ}\right)\\ .\\ .\\ .\\ p\left({\mathrm{\β}}^{\left(i\right)}=k|{\mathrm{\α}}^{\left(i\right)};\mathrm{\θ}\right)\end{array}\right]$

Wherein, h _θ(α) expectation function that softmax classifies is represented, α represents the proper vector that in convolutional neural networks CNN, full articulamentum Fc6 exports, β represents the label corresponding with the proper vector α that articulamentum Fc6 complete in convolutional neural networks CNN exports, p (β=t| α; When θ) representing the proper vector α being input as full articulamentum Fc6 output in convolutional neural networks CNN, label β equals the probability of t, t ∈ 1,2 ..., k, θ represent model parameter and θ ₁, θ ₂..., θ _k∈ R ⁿ⁺¹, softmax Classification Loss function is as follows:

$J\left(\mathrm{\θ}\right)=-\frac{1}{m}\[\underset{i=1}{\overset{m}{\Σ}}{\mathrm{\β}}^{\left(i\right)}{\mathrm{logh}}_{\mathrm{\θ}}\left({\mathrm{\α}}^{\left(i\right)}\right)+(1-{\mathrm{\β}}^{\left(i\right)})log(1-h_{\mathrm{\θ}}({\mathrm{\α}}^{\left(i\right)}))\]$

Wherein, J (θ) represents loss function, and m represents the quantity of car mark areal map sample, h _θ(α) expectation function that softmax classifies is represented, α represents the proper vector that in convolutional neural networks CNN, full articulamentum Fc6 exports, β represents the label corresponding with the proper vector α that articulamentum Fc6 complete in convolutional neural networks CNN exports, and θ represents model parameter.

7. the automobile logo identification method based on convolutional neural networks according to claim 1, is characterized in that: the concrete steps of step (3b) described training convolutional neural networks CNN are as follows:

1st step, in the forward direction stage, gets a sample from sample set, sample is inputted convolutional neural networks CNN and calculate corresponding actual output, in this stage, information through conversion step by step, is sent to convolutional neural networks CNN output layer from convolutional neural networks CNN input layer;

2nd step, in the back-propagation stage, calculates the actual difference exporting the ideal corresponding with sample label and export of convolutional neural networks CNN, by the method for minimization error, and the weights of backpropagation adjustment convolutional neural networks CNN;

3rd step, repeats the operation of the 1st step and the 2nd step, the loss function J (θ)≤0.0001 of the output until convolutional neural networks CNN classifies after layer Softmax7.