CN108460772A - Harassing of advertisement facsimile signal detecting system based on convolutional neural networks and method - Google Patents
Harassing of advertisement facsimile signal detecting system based on convolutional neural networks and method Download PDFInfo
- Publication number
- CN108460772A CN108460772A CN201810150076.0A CN201810150076A CN108460772A CN 108460772 A CN108460772 A CN 108460772A CN 201810150076 A CN201810150076 A CN 201810150076A CN 108460772 A CN108460772 A CN 108460772A
- Authority
- CN
- China
- Prior art keywords
- module
- keyword
- neural network
- facsimile signal
- region
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/155—Segmentation; Edge detection involving morphological operators
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
- G06T2207/10008—Still image; Photographic image from scanner, fax or copier
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of harassing of advertisement facsimile signal detecting system and method based on convolutional neural networks, including keyword region extraction module, the keyword region extraction module are used to determine the keyword suspicious region of facsimile signal to be detected;Neural network Confidence Analysis module, the neural network Confidence Analysis module is connected with the keyword region extraction module, the neural network Confidence Analysis module realizes the classification of facsimile signal for the word of the keyword suspicious region to be identified.The present invention extracts keyword suspicious region by keyword region extraction module, and automatic operating, work efficiency is high;The word of keyword suspicious region is identified by neural network Confidence Analysis module, realizes that the classification of harassing of advertisement fax judges, saves the time, management and control ability is strong so that the present invention has that work efficiency is high, the strong feature of management and control ability.
Description
Technical field
The invention belongs to Image Classfication Technology fields, are disturbed in particular to a kind of advertisement based on convolutional neural networks
Disturb facsimile signal detecting system and method.
Background technology
Universal with Internet technology, the quantity of text image is increasing, text image automatically process be
One important topic of computer application field.The type of text image is various, and layout structure is increasingly sophisticated, includes not only font size
The character area to differ further includes often the elements such as image, table, figure, and the typesetting form of text image is also varied, no
The only rectangle space of a whole page also has the non-rectangle space of a whole page.Automatic business processing is carried out to text image, printed page analysis is an important ring
Section.Printed page analysis is handled text image using computer, and to determine the physical arrangement of text image, image is divided
For the region of the different attributes such as text, image, figure, table, meet text image character recognition, Table recognition, icon-based programming
Etc. different applications needs.Printed page analysis is always the research hotspot in text extracting field, the pre- place as text image
The result of reason process, printed page analysis directly influences the accuracy of subsequent processing.Facsimile signal is important one in text image
Class carries out retrieval to facsimile signal using telephone number, e-mail address for including in facsimile signal etc. and classification obtains
Extensive concern.
Facsimile signal itself has colouring information less, the more low spy different from general optical imagery of spatial resolution
Point.This causes general optical imagery sorting technique to be used directly in the effect in facsimile signal classification problem and bad.By Hubel
Cat visual cortex electrophysiologic study is inspired with Wiesel, it is thus proposed that convolutional neural networks (CNN), Yann Lecun earliest will
CNN is used for Handwritten Digit Recognition.Convolutional neural networks are persistently had an effect in multiple directions in recent years, speech recognition, recognition of face,
There is breakthrough in terms of generic object identification, motion analysis, natural language processing even brain wave analysis.And it is identified in facsimile signal
The application of aspect, convolutional neural networks still belongs to blank.
Gradually deeply, facsimile signal is disturbed as a kind of character image for influence with Internet technology to social life
Disturb presentation ascendant trend.The influence getting worse of the fax of harassing of advertisement in recent years, there is an urgent need to a kind of real-time, high degree of automations
Facsimile signal detection technique.It is reported that the classification of harassing of advertisement facsimile signal at present is relied on and is accomplished manually substantially, exist automatic
Change degree is low, expends the time, the problems such as working efficiency is low, cannot be satisfied supervision needs.
By being analyzed above it is found that existing harassing of advertisement facsimile signal detection method has the following disadvantages:
1, existing harassing of advertisement facsimile signal detection method the degree of automation is low, and working efficiency is low;
2, existing harassing of advertisement facsimile signal detection method expends the time, and prevention ability is poor, cannot meet supervision
It needs.
Invention content
The present invention provides a kind of harassing of advertisement facsimile signal detecting system and method based on convolutional neural networks, can
Effectively solve the problems, such as that existing harassing of advertisement facsimile signal detection method is ineffective, additionally it is possible to solve existing advertisement and disturb
The problem of disturbing facsimile signal detection method management and control energy force difference.
In order to solve problem above, the harassing of advertisement facsimile signal inspection based on convolutional neural networks that the present invention provides a kind of
Examining system and method, technical solution are as follows:
A kind of harassing of advertisement facsimile signal detecting system based on convolutional neural networks, including keyword extracted region mould
Block, the keyword region extraction module are used to determine the keyword suspicious region of facsimile signal to be detected;Neural network confidence
Analysis module is spent, the neural network Confidence Analysis module is connected with the keyword region extraction module, the nerve net
Network Confidence Analysis module is for classifying to the word of the keyword suspicious region.
Such as above-mentioned harassing of advertisement facsimile signal detecting system based on convolutional neural networks, further preferably:It is described
Keyword region extraction module includes binarization block and morphological erosion module;The binarization block is used for the key
Word suspicious region is judged;The morphological erosion module is connected with the binarization block, the morphological erosion module
For corroding to the keyword suspicious region after judgement.
Such as above-mentioned harassing of advertisement facsimile signal detecting system based on convolutional neural networks, further preferably:It is described
Neural network Confidence Analysis module includes input layer, the first convolution module, the second convolution module, third convolution module, the 4th
Convolution module, the 5th convolution module and neural network characteristics grader.
Such as above-mentioned harassing of advertisement facsimile signal detecting system based on convolutional neural networks, further preferably:It is described
Neural network Confidence Analysis module includes 23 layers, and the input layer is the 1st layer of the neural network Confidence Analysis module;
First convolution module is the 2nd to 4 layer of the neural network Confidence Analysis module;Second convolution module is described
The 5th to 7 layer of neural network Confidence Analysis module;The third convolution module is the neural network Confidence Analysis module
The 8th to 11 layer;The Volume Four volume module is the 12nd to 15 layer of the neural network Confidence Analysis module;Described 5th
Convolution module is the 16th to 19 layer of the neural network Confidence Analysis module;The neural network characteristics grader is described
The 20th to 23 layer of neural network Confidence Analysis module.
Such as above-mentioned harassing of advertisement facsimile signal detecting system based on convolutional neural networks, further preferably:It is described
First convolution module, second convolution module, the third convolution module, the Volume Four volume module and the 5th convolution
Module respectively includes convolutional layer and pond layer.
Such as detection method of the above-mentioned harassing of advertisement facsimile signal detecting system based on convolutional neural networks, including it is as follows
Step:
Step 1:Being trained to the convolutional neural networks of the neural network Confidence Analysis module (only need to be for the first time
Training finishes before extracting the keyword suspicious region);
Step 2:Extract the keyword suspicious region;
Step 3:The keyword suspicious region is carried out using the neural network Confidence Analysis module after training
Identification judges.
Such as above-mentioned detection method, further preferably:The neural network characteristics grader is suspicious to the keyword
Region is equipped with keyword confidence level, when the word of the keyword suspicious region is judged as keyword sequence, is then waited for described in judgement
Detection facsimile signal is harassing of advertisement image, is otherwise normal picture.
Such as above-mentioned detection method, further preferably:In step 2, the binarization block is carried out based on OTSU's
Image binaryzation operates, and the facsimile signal to be detected is divided into background and target, and obtain optimal binary-state threshold;The shape
State corrosion module carry out the targeted compression based on morphological erosion, the pixel of the target is compressed, using it is described most
Excellent binary-state threshold judges the keyword suspicious region.
Such as above-mentioned detection method, further preferably:The color of the background sets 255, and the color of the target is set to 0.
Such as above-mentioned detection method, further preferably:In step 3, it is suspicious that the input layer inputs the keyword
Region, first convolution module, second convolution module, the third convolution module, the Volume Four volume module and institute
It states the 5th convolution module successively to handle the keyword suspicious region, the neural network characteristics grader is to the pass
Key word suspicious region is judged.
Analysis is it is found that compared with prior art, the advantages of the present invention are:
1, the harassing of advertisement facsimile signal detecting system provided by the invention based on convolutional neural networks is crucial by being arranged
Word region extraction module, high degree of automation, without human intervention, work efficiency is high;By the way that neural network confidence level point is arranged
Module is analysed, facsimile signal to be detected can quickly be identified, the time is saved, management and control ability is strong so that the present invention has work
Make efficient, the strong feature of management and control ability.
2, the harassing of advertisement facsimile signal detecting system provided by the invention based on convolutional neural networks is by being arranged two-value
Can effectively keyword suspicious region be judged by changing module, by the way that morphological erosion module is arranged, can reduce needs
The number of pixels matched, processing speed is fast, can be suspicious to keyword by being divided to neural network Confidence Analysis module
Region is handled successively, strict logic, so that the present invention has the characteristics that processing speed is fast, strict logic.
3, the harassing of advertisement facsimile signal detecting system provided by the invention based on convolutional neural networks by using based on
The image binaryzation of OTSU operates and the targeted compression based on morphological erosion, is located in advance when extracting keyword suspicious region
Reason, can improve Text region accuracy rate, be handled keyword suspicious region as activation primitive using ReLU, using volume
Product neural network classifies to facsimile signal to be detected, and accuracy is high so that the present invention has the spy efficient, accuracy is high
Point.
Description of the drawings
Fig. 1 is that the harassing of advertisement facsimile signal detecting system based on convolutional neural networks of the present invention constitutes schematic diagram.
Fig. 2 is the image binaryzation operating effect figure based on OTSU of the present invention.
Fig. 3 is the schematic diagram based on morphological erosion of the present invention.
Fig. 4 is the neural network Confidence Analysis module diagram of the present invention.
Fig. 5 is that the facsimile signal to be detected that the present invention carries out carries out the schematic diagram before morphological erosion processing.
Fig. 6 is the morphological erosion treatment effect schematic diagram of the present invention.
Specific implementation mode
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation describes, it is clear that described embodiment is only a part of the embodiment of the present invention, instead of all the embodiments.Base
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts it is all its
His embodiment, shall fall within the protection scope of the present invention.
As shown in Figure 1, the present invention provides a kind of harassing of advertisement facsimile signal detecting system based on convolutional neural networks,
Including keyword region extraction module, keyword region extraction module is for determining the suspicious area of the keyword of facsimile signal to be detected
Domain;Neural network Confidence Analysis module, neural network Confidence Analysis module are connected with keyword region extraction module, nerve
Network Confidence Analysis module is for classifying to the word of keyword suspicious region.
Specifically, the present invention can divide facsimile signal to be detected by the way that keyword region extraction module is arranged
Extraction obtains keyword suspicious region, and neural network Confidence Analysis module, which can be detected keyword suspicious region, to be sentenced
It is disconnected.The present invention is by being arranged keyword region extraction module, and high degree of automation, without human intervention, work efficiency is high;Pass through
Neural network Confidence Analysis module is set, is not necessarily to manual identified, facsimile signal to be detected can quickly be identified, is saved
Time, management and control ability is strong, so that the present invention has, work efficiency is high, the strong feature of management and control ability.
In order to further increase the working efficiency of the present invention, as shown in Figure 1, the keyword region extraction module packet of the present invention
Include binarization block and morphological erosion module;Binarization block is for judging keyword suspicious region;Morphology is rotten
Erosion module is connected with binarization block, and morphological erosion module is for corroding the keyword suspicious region after judgement.This
Invention is few for facsimile signal colouring information, and background is simple, the relatively important feature in pixel geometry position, by the way that binaryzation is arranged
Module can effectively judge keyword suspicious region;By the way that morphological erosion module is arranged, in keyword suspicious region
Under the premise of geometry backbone is constant, it can reduce and need matched number of pixels, shorten processing time, so that the present invention has
There is the characteristics of work efficiency is high.
In order to further increase the management and control ability of the present invention, as shown in Figure 1, the neural network Confidence Analysis mould of the present invention
Block includes input layer, the first convolution module, the second convolution module, third convolution module, Volume Four volume module, the 5th convolution module
With neural network characteristics grader.The present invention, can be to keyword by being divided to neural network Confidence Analysis module
Suspicious region is handled successively, strict logic, can effectively avoid there is a situation where keyword suspicious region judge by accident, to
So that the present invention has the characteristics that management and control ability is strong.
In order to which the neural network Confidence Analysis module to the present invention carries out work distribution, as shown in Figure 1, the god of the present invention
It include 23 layers through network Confidence Analysis module, input layer is the 1st layer of neural network Confidence Analysis module;First volume product module
Block is the 2nd to 4 layer of neural network Confidence Analysis module;Second convolution module is the of neural network Confidence Analysis module
5 to 7 layers;Third convolution module is the 8th to 11 layer of neural network Confidence Analysis module;Volume Four volume module is neural network
The 12nd to 15 layer of Confidence Analysis module;5th convolution module is the 16th to 19 layer of neural network Confidence Analysis module;
Neural network characteristics grader is the 20th to 23 layer of neural network Confidence Analysis module.The present invention is by setting neural network
Reliability Analysis module carries out classifying rationally, is capable of the work of every layer of reasonable distribution neural network Confidence Analysis module, improves work
Make efficiency, so that the present invention has the characteristics that work efficiency is high.
In order to simplify the present invention parameter operation, as shown in Figure 1, the present invention the first convolution module, the second convolution module,
Third convolution module, Volume Four volume module and the 5th convolution module respectively include convolutional layer and pond layer.The present invention passes through every
A convolution module setting convolutional layer and pond layer, enormously simplify the complexity of operational model, reduce the parameter of model, to
So that the present invention has the characteristics that parameter simple operation.
As shown in Figures 1 to 6, the harassing of advertisement facsimile signal inspection based on convolutional neural networks that the present invention also provides a kind of
The detection method of examining system, includes the following steps:
Step 1:The convolutional neural networks of neural network Confidence Analysis module are trained.
1.1 seek initial training model:
A data set for including 1000 Chinese characters is chosen, each Chinese character there are 256 images, and Image Acquisition is from different hands
Text is write, and is normalized, initial training model is obtained.
The second training of 1.2 models:
On the basis of 1.1, the mode of transfer learning is taken, carries out second training.Manually determining harassing of advertisement is passed
True image is syncopated as A, B, C, multiple subgraphs such as D manually from caption position, and each subgraph is a character.For each
Subgraph, 45 degree clockwise rotate 3 times, and do flip horizontal to each position, obtain 8 × 2 images.At random [224,
386] size is selected between, and every image resize (adjustment) is arrived into this size, on it crop (acquisition) one
Size is the region of (224,224) as input.
1.3 obtain complete convolutional neural networks training set:
The training set of 1.2 compositions is upset at random, calculate the mean value of training set image and carries out averaging operation, after processing
Result be convolutional neural networks complete training set.
Convolutional neural networks in the training process, need to carry out average value processing to training set, and in test process and are
System deployment process need not carry out average value processing.The batch of network training is sized to 64, using SGD (Stochastic
The abbreviation of gradient descent, stochastic gradient descent) algorithm.Momentum is set as 0.9, and maximum iteration is 100,000 times,
Initial learning rate is 0.05, every 10,000 decaying 0.0005.The L2 regular coefficients of L2 regularizations are 0.1, reduce over-fitting and use
Dropout modes, probability are set as 0.5.In network description write-in prototxt files, training parameter write-in
In solver.prototxt files, caffe train orders is called to be trained, the .caffemodel preserved after training
File includes network training parametric results.
Step 2:Extract keyword suspicious region.
The 2.1 image binaryzation operations based on OTSU:
Binarization block carries out the operation of the image binaryzation based on OTSU, using maximum variance between clusters (OTSU) to be checked
It surveys and chooses an optimal binary-state threshold in facsimile signal, facsimile signal to be detected is divided into background and target, color of object is set to 0
(black), background color sets 255 (white), to reach binaryzation, enables to the gap of the same category between target and background most
It is small, different classes of disparity.Wherein, optimal binary-state threshold acquiring method is:
If f (x, y) indicates that the gray scale at pixel (x, y), g (x, y) indicate 3 × 3 neighborhood average gray of (x, y),
Then g (x, y) is represented by:
If the size of facsimile signal to be detected is M × N, 0≤x+m≤M-1,0≤y+n≤N-1 enable i=f (x, y) and j=
G (x, y) composition of vector (i, j), if CijIndicate the number that (i, j) occurs, then the probability that (i, j) occurs is:
If facsimile signal to be detected is divided into target and background by binary-state threshold (s, t), then this two parts occurs general
Rate is:
Gray average vector is:
Mean value vector on two-dimensional histogram is:
Inter-class variance is:
For background and target, the similarity degree of this two class is described with inter-class variance δ, δ is bigger, the difference of background and target
It is bigger, it is on the contrary then smaller.If background area is classified as target area or target area is classified as background area, inter-class variance δ
It will reduce.Therefore, when the inter-class variance of target and background is maximum and when variance within clusters minimum, obtained segmentation result effect
Fruit is best.Corresponding (s, t) value is optimal binary-state threshold when δ is maximized.Treatment effect is as shown in Figure 2.
Using the image binaryzation operation based on OTSU (maximum variance between clusters), phase between target and background is enabled to
Generic gap is minimum, different classes of disparity, this method when inter-class variance maximum in facsimile signal to be detected
Best results.Otsu is carried out based on the histogram of facsimile signal to be detected, and operation is simple and fast, therefore Otsu divides efficiency
Height is widely used.
2.2 targeted compressions based on morphological erosion:
Morphological erosion module carries out the targeted compression based on morphological erosion, is compressed to the pixel of target, form
Learning the mathematic(al) representation corroded is:
That is moving structure B then preserves the position if the intersection of fruit structure B and structure A are fully located in the region of structure A
Point, all location points for meeting condition constitute the result that structure A is corroded by structure B.When B is structure shown in Fig. 3, A can be risen
To compression effectiveness, handling principle is as shown in figure 3, as shown in figure 5, treatment effect such as Fig. 6 institutes before carrying out morphological erosion processing
Show.Morphological erosion can extract the connected region skeleton of target under the premise of target geometry backbone is constant, reduce needs
The number of pixels matched shortens to reduce the calculation amount of matching operation and calculates the time.
2.3 extraction keyword suspicious regions:
If the set of keywords identified is needed to be combined into { A, B, C, D }, random selection one is to be checked from each set element
Facsimile signal is surveyed as region template, after step 1.1 and step 1.2 processing, all templates are asked and obtain keyword can
Region is doubted, the segmentation to facsimile signal to be detected can be completed, extraction keyword suspicious region is sent into neural network confidence level point
Analysis module is classified.Facsimile signal belongs to one kind of character image, has colouring information few, background is simple, pixel geometry position
Confidence ceases relatively important feature.Text region accuracy rate can be improved by carrying out step 1.1 and step 1.2.
Step 3:Judgement is identified to keyword suspicious region.
Neural network Confidence Analysis module carries out keyword suspicious region processing using ReLU as activation primitive, uses
SOFTMAX algorithms classify to facsimile signal to be detected.As shown in figure 4, identification judges that flow is:INPUT→[CONV×m
→ POOL] × 5 → FC × 3 → SOFTMAX, wherein INPUT is input layer, and CONV is convolutional layer, and ReLU is nonlinear activation letter
Number, POOL are pond layer, and FC is full articulamentum.
3.1 input keyword suspicious regions:
The 1st layer of neural network Confidence Analysis module is input layer, and input layer is for inputting keyword suspicious region.
3.2 pairs of keyword suspicious regions are handled:
The 2nd, 3,4 layer of neural network Confidence Analysis module is the first convolution module, wherein 2,3 layers are convolutional layer, often
Layer has the convolution kernels of 64 3 × 3 sizes, can ensure up and down, left and right, in these direction concepts the minimum perception visual field it is big
Small, step value 1, convolution operation ensures spatial resolution by the way of zero padding.4th layer is 2 × 2 sizes, and step value is 2
Pond layer, in such a way that maximization is down-sampled.
The 5th, 6,7 layer of neural network Confidence Analysis module is the second convolution module, and the 5th, 6 layer of convolution nuclear volume becomes
128, it can make up since down-sampled caused spatial resolution reduces, the 7th layer of pond layer is the same.
The 8th, 9,10,11 layer of neural network Confidence Analysis module is third convolution module, and the 8th, 9,10 layer is convolution
Layer, it is 256 that convolution quantity, which increases, increases by one layer of convolutional layer compared to before, 11th layer pond layer is the same.
The 12nd, 13,14,15 layer of neural network Confidence Analysis module be Volume Four volume module, the 16th, 17,18,19 layer
For the 5th convolution module.Volume Four volume module and the 5th convolution module structure are identical as third convolution module.
In the convolutional neural networks of neural network Confidence Analysis module, a neuron only connects with partial nerve member
It connects.In a convolutional layer of convolutional neural networks, including several characteristic planes, each characteristic plane is by some rectangular arrangeds
Neuron composition, the neuron of same characteristic plane shares convolution kernel.Convolution kernel is initialized in the form of random matrix, in net
Study is obtained rational weights by convolution kernel in the training process of network.Shared convolution nuclear energy enough reduces the company between each layer of network
It connects, while reducing the risk of over-fitting again.Convolutional neural networks reduce in such a way that local sensing open country and convolution kernel are shared
Number of parameters, it is only necessary to the part of facsimile signal to be detected be perceived, then carried out the information of part in higher comprehensive
It closes, to obtain the global information of facsimile signal to be detected, and a certain Partial Feature of facsimile signal to be detected can be also used in
On its another part, for all positions on facsimile signal to be detected, same learning characteristic can be used.
3.3 output facsimile signal recognition results to be detected:
The 20th, 21,22,23 layer of neural network Confidence Analysis module is neural network characteristics grader, the 20th layer and the
21 layers contain 4096 parameters units, treated for placing step 2.2 keyword suspicious region, and the 22nd layer is equipped with key
Word confidence level, the 23rd layer exports facsimile signal testing result to be detected using softmax algorithms.The word of keyword suspicious region
When being judged as keyword sequence by the 22nd layer, for harassing of advertisement image, otherwise the 23rd layer then judges the facsimile signal to be detected
For normal picture.
Neural network Confidence Analysis module with the small convolution kernel of multilayer (such as 33 × 3) replace the big convolution kernel of single layer (1 7 ×
7).With the increase of depth, discrimination improves significantly, and work efficiency is high, and management and control ability is strong.
Analysis is it is found that compared with prior art, the advantages of the present invention are:
1, the harassing of advertisement facsimile signal detecting system provided by the invention based on convolutional neural networks is crucial by being arranged
Word region extraction module, high degree of automation, without human intervention, work efficiency is high;By the way that neural network confidence level point is arranged
Module is analysed, facsimile signal to be detected can quickly be identified, the time is saved, management and control ability is strong so that the present invention has work
Make efficient, the strong feature of management and control ability.
2, the harassing of advertisement facsimile signal detecting system provided by the invention based on convolutional neural networks is by being arranged two-value
Can effectively keyword suspicious region be judged by changing module, by the way that morphological erosion module is arranged, can reduce needs
The number of pixels matched, processing speed is fast, can be suspicious to keyword by being divided to neural network Confidence Analysis module
Region is handled successively, strict logic, so that the present invention has the characteristics that processing speed is fast, strict logic.
3, the harassing of advertisement facsimile signal detecting system provided by the invention based on convolutional neural networks by using based on
The image binaryzation of OTSU operates and the targeted compression based on morphological erosion, is located in advance when extracting keyword suspicious region
Reason, can improve Text region accuracy rate, be handled keyword suspicious region as activation primitive using ReLU, using volume
Product neural network classifies to facsimile signal to be detected, and accuracy is high so that the present invention has the spy efficient, accuracy is high
Point.
As known by the technical knowledge, the present invention can pass through the embodiment party of other essence without departing from its spirit or essential feature
Case is realized.Therefore, embodiment disclosed above, all things considered are all merely illustrative, not the only.Institute
Have within the scope of the present invention or in the change being equal in the scope of the present invention and includes by the present invention.
Claims (10)
1. a kind of harassing of advertisement facsimile signal detecting system based on convolutional neural networks, which is characterized in that including:
Keyword region extraction module, the keyword region extraction module are used to determine that the keyword of facsimile signal to be detected can
Doubt region;
Neural network Confidence Analysis module, the neural network Confidence Analysis module and the keyword region extraction module
It is connected, the neural network Confidence Analysis module is for classifying to the word of the keyword suspicious region.
2. the harassing of advertisement facsimile signal detecting system according to claim 1 based on convolutional neural networks, feature exist
In:
The keyword region extraction module includes binarization block and morphological erosion module;The binarization block for pair
The keyword suspicious region is judged;The morphological erosion module is connected with the binarization block, the morphology
Corrosion module is for corroding the keyword suspicious region after judgement.
3. the harassing of advertisement facsimile signal detecting system according to claim 1 based on convolutional neural networks, feature exist
In:
The neural network Confidence Analysis module includes input layer, the first convolution module, the second convolution module, third convolution mould
Block, Volume Four volume module, the 5th convolution module and neural network characteristics grader.
4. the harassing of advertisement facsimile signal detecting system according to claim 3 based on convolutional neural networks, feature exist
In:
The neural network Confidence Analysis module includes 23 layers, and the input layer is the neural network Confidence Analysis module
The 1st layer;First convolution module is the 2nd to 4 layer of the neural network Confidence Analysis module;The volume Two product module
Block is the 5th to 7 layer of the neural network Confidence Analysis module;The third convolution module is the neural network confidence level
The 8th to 11 layer of analysis module;The Volume Four volume module is the 12nd to 15 layer of the neural network Confidence Analysis module;
5th convolution module is the 16th to 19 layer of the neural network Confidence Analysis module;The neural network characteristics classification
Device is the 20th to 23 layer of the neural network Confidence Analysis module.
5. the harassing of advertisement facsimile signal detecting system according to claim 4 based on convolutional neural networks, feature exist
In:
First convolution module, second convolution module, the third convolution module, the Volume Four volume module and described
5th convolution module respectively includes convolutional layer and pond layer.
6. the harassing of advertisement facsimile signal detection system based on convolutional neural networks according to claim 1 to 5 any one
The detection method of system, which is characterized in that include the following steps:
Step 1:The convolutional neural networks of the neural network Confidence Analysis module are trained;
Step 2:Extract the keyword suspicious region;
Step 3:The keyword suspicious region is identified using the neural network Confidence Analysis module after training
Judge.
7. detection method according to claim 6, it is characterised in that:
The neural network characteristics grader is equipped with keyword confidence level to the keyword suspicious region, and the keyword is suspicious
When the word in region is judged as keyword sequence, then the facsimile signal to be detected is judged for harassing of advertisement image, otherwise for just
Normal image.
8. detection method according to claim 7, it is characterised in that:
In step 2, the binarization block carries out the operation of the image binaryzation based on OTSU, by the facsimile chart to be detected
As being divided into background and target, and obtain optimal binary-state threshold;The morphological erosion module is carried out based on morphological erosion
Targeted compression compresses the pixel of the target, using the optimal binary-state threshold to the keyword suspicious region
Judged.
9. the detection side of the harassing of advertisement facsimile signal detecting system according to claim 8 based on convolutional neural networks
Method, it is characterised in that:
The color of the background sets 255, and the color of the target is set to 0.
10. the detection side of the harassing of advertisement facsimile signal detecting system according to claim 9 based on convolutional neural networks
Method, it is characterised in that:
In step 3, the input layer inputs the keyword suspicious region, first convolution module, second convolution
Module, the third convolution module, the Volume Four volume module and the 5th convolution module are suspicious to the keyword successively
Region is handled, and the neural network characteristics grader judges the keyword suspicious region.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810150076.0A CN108460772B (en) | 2018-02-13 | 2018-02-13 | Advertisement harassment fax image detection system and method based on convolutional neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810150076.0A CN108460772B (en) | 2018-02-13 | 2018-02-13 | Advertisement harassment fax image detection system and method based on convolutional neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108460772A true CN108460772A (en) | 2018-08-28 |
CN108460772B CN108460772B (en) | 2022-05-17 |
Family
ID=63216529
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810150076.0A Active CN108460772B (en) | 2018-02-13 | 2018-02-13 | Advertisement harassment fax image detection system and method based on convolutional neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108460772B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111144399A (en) * | 2018-11-06 | 2020-05-12 | 富士通株式会社 | Apparatus and method for processing image |
CN112561055A (en) * | 2020-12-09 | 2021-03-26 | 北京交通大学 | Electromagnetic disturbance identification method based on bilinear time-frequency analysis and convolutional neural network |
Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150036920A1 (en) * | 2013-07-31 | 2015-02-05 | Fujitsu Limited | Convolutional-neural-network-based classifier and classifying method and training methods for the same |
CN105574524A (en) * | 2015-12-11 | 2016-05-11 | 北京大学 | Cartoon image page identification method and system based on dialogue and storyboard united identification |
US20160283841A1 (en) * | 2015-03-27 | 2016-09-29 | Google Inc. | Convolutional neural networks |
CN106096602A (en) * | 2016-06-21 | 2016-11-09 | 苏州大学 | Chinese license plate recognition method based on convolutional neural network |
US9501724B1 (en) * | 2015-06-09 | 2016-11-22 | Adobe Systems Incorporated | Font recognition and font similarity learning using a deep neural network |
CN106156777A (en) * | 2015-04-23 | 2016-11-23 | 华中科技大学 | Textual image detection method and device |
CN106650721A (en) * | 2016-12-28 | 2017-05-10 | 吴晓军 | Industrial character identification method based on convolution neural network |
US20170206431A1 (en) * | 2016-01-20 | 2017-07-20 | Microsoft Technology Licensing, Llc | Object detection and classification in images |
CN107437100A (en) * | 2017-08-08 | 2017-12-05 | 重庆邮电大学 | A kind of picture position Forecasting Methodology based on the association study of cross-module state |
CN107578060A (en) * | 2017-08-14 | 2018-01-12 | 电子科技大学 | A kind of deep neural network based on discriminant region is used for the method for vegetable image classification |
WO2018010434A1 (en) * | 2016-07-13 | 2018-01-18 | 华为技术有限公司 | Image classification method and device |
-
2018
- 2018-02-13 CN CN201810150076.0A patent/CN108460772B/en active Active
Patent Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150036920A1 (en) * | 2013-07-31 | 2015-02-05 | Fujitsu Limited | Convolutional-neural-network-based classifier and classifying method and training methods for the same |
US20160283841A1 (en) * | 2015-03-27 | 2016-09-29 | Google Inc. | Convolutional neural networks |
CN106156777A (en) * | 2015-04-23 | 2016-11-23 | 华中科技大学 | Textual image detection method and device |
US9501724B1 (en) * | 2015-06-09 | 2016-11-22 | Adobe Systems Incorporated | Font recognition and font similarity learning using a deep neural network |
CN105574524A (en) * | 2015-12-11 | 2016-05-11 | 北京大学 | Cartoon image page identification method and system based on dialogue and storyboard united identification |
US20170206431A1 (en) * | 2016-01-20 | 2017-07-20 | Microsoft Technology Licensing, Llc | Object detection and classification in images |
CN106096602A (en) * | 2016-06-21 | 2016-11-09 | 苏州大学 | Chinese license plate recognition method based on convolutional neural network |
WO2018010434A1 (en) * | 2016-07-13 | 2018-01-18 | 华为技术有限公司 | Image classification method and device |
CN106650721A (en) * | 2016-12-28 | 2017-05-10 | 吴晓军 | Industrial character identification method based on convolution neural network |
CN107437100A (en) * | 2017-08-08 | 2017-12-05 | 重庆邮电大学 | A kind of picture position Forecasting Methodology based on the association study of cross-module state |
CN107578060A (en) * | 2017-08-14 | 2018-01-12 | 电子科技大学 | A kind of deep neural network based on discriminant region is used for the method for vegetable image classification |
Non-Patent Citations (6)
Title |
---|
CHRIS TENSMEYER等: "Document Image Binarization with Fully Convolutional Neural Networks", 《ARXIV.ORG/PDF/1708.03276.PDF》 * |
MICHAEL FIRE等: "Exploring Online Ad Images Using a Deep Convolutional Neural Network Approach", 《COMPUTER VISION AND PATTERN RECOGNITION》 * |
XIAOFAN LIN: "Towards Accurate Binary Convolutional Neural Network", 《31ST CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS(NIPS2017)》 * |
关鑫: "卷积神经网络在手写体识别的应用", 《电子技术与软件工程》 * |
张红等: "基于卷积神经网络的手写数字识别算法", 《电子技术与软件工程》 * |
郭晓宇等: "基于连通区域的传真图像版面分割与分类算法", 《计算机应用研究》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111144399A (en) * | 2018-11-06 | 2020-05-12 | 富士通株式会社 | Apparatus and method for processing image |
CN111144399B (en) * | 2018-11-06 | 2024-03-05 | 富士通株式会社 | Apparatus and method for processing image |
CN112561055A (en) * | 2020-12-09 | 2021-03-26 | 北京交通大学 | Electromagnetic disturbance identification method based on bilinear time-frequency analysis and convolutional neural network |
CN112561055B (en) * | 2020-12-09 | 2023-11-07 | 北京交通大学 | Electromagnetic disturbance identification method based on bilinear time-frequency analysis and convolutional neural network |
Also Published As
Publication number | Publication date |
---|---|
CN108460772B (en) | 2022-05-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wei et al. | Inverse discriminative networks for handwritten signature verification | |
Srikantan et al. | Gradient-based contour encoding for character recognition | |
Yuan et al. | Bag-of-words and object-based classification for cloud extraction from satellite imagery | |
CN109002766A (en) | A kind of expression recognition method and device | |
CN109711448A (en) | Based on the plant image fine grit classification method for differentiating key field and deep learning | |
CN106228166B (en) | The recognition methods of character picture | |
Fazilov et al. | State of the art of writer identification | |
Raj et al. | Helmet violation processing using deep learning | |
CN109886161A (en) | A kind of road traffic index identification method based on possibility cluster and convolutional neural networks | |
CN111860309A (en) | Face recognition method and system | |
Hebbale et al. | Real time COVID-19 facemask detection using deep learning | |
CN107220598A (en) | Iris Texture Classification based on deep learning feature and Fisher Vector encoding models | |
Karunarathne et al. | Recognizing ancient sinhala inscription characters using neural network technologies | |
CN106203448A (en) | A kind of scene classification method based on Nonlinear Scale Space Theory | |
CN108460772A (en) | Harassing of advertisement facsimile signal detecting system based on convolutional neural networks and method | |
CN114359917A (en) | Handwritten Chinese character detection and recognition and font evaluation method | |
CN114782979A (en) | Training method and device for pedestrian re-recognition model, storage medium and terminal | |
CN114581928A (en) | Form identification method and system | |
CN114386413A (en) | Handling digitized handwriting | |
CN109741351A (en) | A kind of classification responsive type edge detection method based on deep learning | |
CN105844299B (en) | A kind of image classification method based on bag of words | |
Jingyi et al. | Classification of images by using TensorFlow | |
CN110633666A (en) | Gesture track recognition method based on finger color patches | |
CN116958615A (en) | Picture identification method, device, equipment and medium | |
CN111325270B (en) | Dongba text recognition method based on template matching and BP neural network |
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 |