CN107346629A - A kind of intelligent blind reading method and intelligent blind reader system - Google Patents
A kind of intelligent blind reading method and intelligent blind reader system Download PDFInfo
- Publication number
- CN107346629A CN107346629A CN201710739516.1A CN201710739516A CN107346629A CN 107346629 A CN107346629 A CN 107346629A CN 201710739516 A CN201710739516 A CN 201710739516A CN 107346629 A CN107346629 A CN 107346629A
- Authority
- CN
- China
- Prior art keywords
- text
- module
- usb
- intelligent blind
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 40
- 238000004891 communication Methods 0.000 claims abstract description 31
- 238000012545 processing Methods 0.000 claims abstract description 30
- 230000011218 segmentation Effects 0.000 claims abstract description 23
- 238000012549 training Methods 0.000 claims description 21
- 238000013527 convolutional neural network Methods 0.000 claims description 13
- 230000015572 biosynthetic process Effects 0.000 claims description 9
- 238000003786 synthesis reaction Methods 0.000 claims description 9
- 101000941170 Homo sapiens U6 snRNA phosphodiesterase 1 Proteins 0.000 claims description 6
- 102100031314 U6 snRNA phosphodiesterase 1 Human genes 0.000 claims description 6
- 238000004458 analytical method Methods 0.000 claims description 5
- 238000011144 upstream manufacturing Methods 0.000 claims description 5
- 230000008859 change Effects 0.000 claims description 4
- 238000013135 deep learning Methods 0.000 claims description 3
- -1 USB3 and USB2 Proteins 0.000 claims description 2
- 230000008901 benefit Effects 0.000 abstract description 6
- 230000006870 function Effects 0.000 description 16
- 239000011159 matrix material Substances 0.000 description 12
- 230000008569 process Effects 0.000 description 12
- 230000000694 effects Effects 0.000 description 11
- 238000010586 diagram Methods 0.000 description 10
- 238000005516 engineering process Methods 0.000 description 10
- 238000012360 testing method Methods 0.000 description 8
- 230000005540 biological transmission Effects 0.000 description 7
- 238000013461 design Methods 0.000 description 7
- 238000001514 detection method Methods 0.000 description 7
- 238000004422 calculation algorithm Methods 0.000 description 5
- 230000000644 propagated effect Effects 0.000 description 5
- 230000004913 activation Effects 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 4
- 238000012937 correction Methods 0.000 description 4
- 210000002569 neuron Anatomy 0.000 description 4
- 230000006835 compression Effects 0.000 description 3
- 238000007906 compression Methods 0.000 description 3
- 230000007246 mechanism Effects 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000005265 energy consumption Methods 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 238000009499 grossing Methods 0.000 description 2
- 238000010191 image analysis Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 230000023886 lateral inhibition Effects 0.000 description 2
- 230000015654 memory Effects 0.000 description 2
- 238000010295 mobile communication Methods 0.000 description 2
- 238000000465 moulding Methods 0.000 description 2
- 230000006855 networking Effects 0.000 description 2
- 230000007935 neutral effect Effects 0.000 description 2
- 230000002093 peripheral effect Effects 0.000 description 2
- 238000011176 pooling Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 201000009487 Amblyopia Diseases 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 239000003086 colorant Substances 0.000 description 1
- 238000004040 coloring Methods 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000003708 edge detection Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000001771 impaired effect Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 230000003278 mimic effect Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 210000000653 nervous system Anatomy 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 238000004806 packaging method and process Methods 0.000 description 1
- 238000007639 printing Methods 0.000 description 1
- 230000008707 rearrangement Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012772 sequence design Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- 208000029257 vision disease Diseases 0.000 description 1
- 230000001755 vocal effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G09—EDUCATION; CRYPTOGRAPHY; DISPLAY; ADVERTISING; SEALS
- G09B—EDUCATIONAL OR DEMONSTRATION APPLIANCES; APPLIANCES FOR TEACHING, OR COMMUNICATING WITH, THE BLIND, DEAF OR MUTE; MODELS; PLANETARIA; GLOBES; MAPS; DIAGRAMS
- G09B21/00—Teaching, or communicating with, the blind, deaf or mute
- G09B21/001—Teaching or communicating with blind persons
- G09B21/006—Teaching or communicating with blind persons using audible presentation of the information
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- General Health & Medical Sciences (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- Educational Administration (AREA)
- Educational Technology (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Character Discrimination (AREA)
Abstract
The invention discloses a kind of intelligent blind reading method and intelligent blind reader system, and the image information of text is gathered by image capture module, the image information of text is uploaded into Cloud Server by communication module by USB diverter modules；The image information of Cloud Server collection is pre-processed, Text segmentation, Text region, and recognition result is sent into USB diverter modules with text formatting；USB diverter modules are retransmited to central control module, the recognition result of text formatting is changed into voice messaging by central control module by the Audio Processing Unit of voice broadcast module, bluetooth unit is sent, by the earphone or loudspeaker of bluetooth unit, by voice broadcast to reader.Compared with prior art, there is the identification for realizing the eight kinds of fonts commonly used in printed Chinese character, realize that handwritten Chinese character body identifies, reading efficiency is high, reads the advantages of accuracy rate is high.
Description
Technical field
The present invention relates to a kind of reader system, particularly a kind of intelligent blind reader recognition methods and intelligent blind are read
Read device system.
Background technology
No. 282 commentary updated according to the World Health Organization (WHO) in August, 2014, in the range of the world today
Visually impaired number is about 2.85 hundred million, wherein 39,000,000 people suffer from blind disease, 2.46 hundred million people are amblyopia, and developing country regards
Power sufferer's number accounts for global 90%.In addition, China carried out the Second China National Sample Survey on Disability in 2006,
Calculated according to investigation result, the sum of all kinds of disabled persons in China in 2006 is 82,960,000 people, wherein 12,330,000 people with visual disabilities.Root
The whole nation is accounted for according to the 6th national census China total number of people, and the Second China National Sample Survey on Disability China disabled person always
The ratio of population and all kinds of disabled persons account for the ratio of disabled person's total number of persons, calculate 2010 year end China disabled person total numbers of persons 85,020,000
People, wherein number with visual disabilities is about 12,630,000 people.
As can be seen that having, the crowd of vision disorder is very huge, and as the improvement of people's living standards, reading increasingly
As a kind of strong demand, while with the development of modern science and technology, our life enters an information content increasing, information
The circulation way diversified New Times, this is the epoch for being properly termed as " Information barrier-free ", exactly makes no matter abled person goes back
It is disabled person, no matter young man or the elderly can benefit from information technology, and anyone can put down under any circumstance
Etc. ground, easily, obtain information without barrier, utilize information." Information barrier-free " is also known as " information accessibility ", and its target is will
Make whole society all people, including physical disabilities, the elderly and children, the chance with impartial acquisition or use information.It is modern
The achievement of scientific and technological development of civilization, it should benefit everyone.Blind person is a urgent need colony to be aided among these, and they also have
The serious hope urgent to external information.But the mode of blind person's study now is substantially armrest and refers to touch braille to be read a book,
Not only reading efficiency is low for this mode, and the accuracy of reading content is not also high, and this study to blind person causes very great
Divide difficulty.Existing intelligent blind reader can not preferably realize eight kinds of fonts (Song typeface, the patterns commonly used in printed Chinese character
Body, lishu, black matrix, children's circle, Chinese-language row pattern, the new Wei of Chinese, the easypro body of Chinese) identification and realize that handwritten Chinese character body identifies.
Therefore, existing reading aid for blindmen deposit can not realize in printed Chinese character commonly use the Song typeface, regular script, lishu, black matrix,
The new Wei of children's circle, Chinese-language row pattern, Chinese, the identification of the easypro body of Chinese, can not realize that handwritten Chinese character body identifies, reading efficiency is low, reads
The shortcomings that accuracy rate is low.
The content of the invention
It is an object of the present invention to provide a kind of intelligent blind reader system, the present invention, which has, realizes the block letter Chinese
The identification of the Song typeface, regular script, lishu, black matrix, children's circle, Chinese-language row pattern, the new Wei of Chinese, the easypro body of Chinese commonly used in word, is realized hand-written
Chinese Character Font Recognition, reading efficiency is high, reads the advantages of accuracy rate is high.
Technical scheme：A kind of intelligent blind reading method, the image of text is gathered by image capture module
Information, the image information of text is uploaded to by Cloud Server by communication module by USB diverter modules；The figure of Cloud Server collection
As information pre-processed, Text segmentation, Text region, and recognition result is sent to USB diverter modules with text formatting；
USB diverter modules retransmit will by the Audio Processing Unit of voice broadcast module to central control module, central control module
The recognition result of text formatting changes into voice messaging, send bluetooth unit, by the earphone or loudspeaker of bluetooth unit, by voice
Report to reader.
Foregoing intelligent blind reading method, the Cloud Server be by image analysis system complete picture pre-process,
Text segmentation and Text region.
Foregoing intelligent blind reading method, described Text region, it is to be completed using word training pattern, uses depth
Training pattern of the convolutional neural networks as single character recognition in study.
The intelligent blind reader system of foregoing intelligent blind reading method, including central control module, center control
Molding block is connected to voice broadcast module and USB diverter modules, and USB diverter modules are connected by communication module and Cloud Server
Connect, USB diverter modules are also associated with image capture module.
Foregoing intelligent blind reader system, the voice broadcast module include Audio Processing Unit, speech processes list
Member is connected with bluetooth unit.
Foregoing intelligent blind reader system, the phonetic synthesis unit include serial ports transmitting-receiving, phonetic synthesis and voice
Outlet line, the bluetooth unit include power amplifier and loudspeaker；The input and central control module of serial ports transmitting-receiving
UART3 connections, the output end of serial ports transmitting-receiving by phonetic synthesis and voice output connection, the output of voice output circuit according to
It is secondary to be connected by the power amplifier of bluetooth unit with loudspeaker.
Foregoing intelligent blind reader system, the central control module include main control chip, and main control chip passes through slow
Deposit, processing and control element (PCE) is connected with UART3 and UART2 respectively, UART2 is connected with reserved serial ports, and caching, processing and control element (PCE) are also
USB.HOST 2.0 is connected with, USB.HOST 2.0 is connected with the upstream port of USB diverter modules.
Foregoing intelligent blind reader system, the USB diverter modules include multiport transmitting-receiving control unit, multiport
Transmitting-receiving control unit is connected to upstream port, exchanges, changes, caching, processing, USB1, USB3 and USB2, USB1 and image
The spare interface of the camera connection of acquisition module, USB3 and image capture module connects, the USB interface of USB2 and communication module
Connection.
Foregoing intelligent blind reader system, the communication module include 4G communication chips, and 4G communication chips connect respectively
Radio-frequency antenna, SIM card and USB interface are connected to, radio-frequency antenna is connected with Cloud Server.
Beneficial effects of the present invention：Compared with prior art, present invention incorporates machine vision, Digital Image Processing, depth
Spend study, the cutting edge technology such as computer network, can not only realize the identification of printed Chinese character, but also can realize the Song typeface,
Eight kinds of Character Font Recognitions such as regular script, lishu, black matrix.Further, it is also possible to realize the identification of handwritten Chinese character body.Traditional blind person is broken
Read books need with finger touch blind person limitation, have realize in printed Chinese character commonly use the Song typeface, regular script, lishu,
The identification of black matrix, children's circle, Chinese-language row pattern, the new Wei of Chinese, the easypro body of Chinese, realizes that handwritten Chinese character body identifies, reading efficiency is high, reads
The advantages of accuracy rate is high.
Brief description of the drawings
Fig. 1 is the theory diagram of present system；
Fig. 2 is the hardcore structured flowchart in Fig. 1；
Fig. 3 is the fundamental diagram of voice broadcast module in Fig. 1；
Fig. 4 is the fundamental diagram of communication module in Fig. 1；
Fig. 5 is the fundamental diagram of USB diverter modules in Fig. 1；
Fig. 6 is the operational flow diagram of the present invention；
Fig. 7 is the word graph of the corresponding generation of various fonts；
Fig. 8 is CNN implementation model simplification figures；
Fig. 9 is verbal model training network illustraton of model；
Figure 10 is SoftmaxWithLoss layer schematic diagrames；
Figure 11 is the graph of a relation of accuracy and iterations.
Embodiment
The present invention is further illustrated with reference to the accompanying drawings and examples, but be not intended as to the present invention limit according to
According to.
Embodiment.A kind of intelligent blind reading method and intelligent blind reader system, form as shown in Fig. 1~11, one
Kind intelligent blind reading method, the image information of text is gathered by image capture module, passes through the mould that communicates by USB diverter modules
The image information of text is uploaded to Cloud Server by block；The image information of Cloud Server collection is pre-processed, Text segmentation, text
Word is identified, and recognition result is sent into USB diverter modules with text formatting；USB diverter modules retransmit to center and control mould
Block, the recognition result of text formatting is changed into voice by central control module by the Audio Processing Unit of voice broadcast module to be believed
Breath, send bluetooth unit, by the earphone or loudspeaker of bluetooth unit, by voice broadcast to reader.
The Cloud Server is to complete picture pretreatment, Text segmentation and Text region by image analysis system.
Described Text region, it is to be completed using word training pattern, is made using the convolutional neural networks in deep learning
For the training pattern of single character recognition.
The intelligent blind reader system of described intelligent blind reading method, including central control module, center control
Molding block is connected to voice broadcast module and USB diverter modules, and USB diverter modules are connected by communication module and Cloud Server
Connect, USB diverter modules are also associated with image capture module.
The voice broadcast module includes Audio Processing Unit, and Audio Processing Unit is connected with bluetooth unit.
The phonetic synthesis unit, which includes serial ports transmitting-receiving, phonetic synthesis and voice output circuit, the bluetooth unit, to be included
Power amplifier and loudspeaker；The UART3 connections of the input and central control module of serial ports transmitting-receiving, the output end of serial ports transmitting-receiving are led to
Cross phonetic synthesis and voice output connection, the output of voice output circuit pass sequentially through bluetooth unit power amplifier and
Loudspeaker connect.
The central control module includes main control chip, main control chip by caching, processing and control element (PCE) respectively with UART3
Being connected with UART2, UART2 is connected with reserved serial ports, and caching, processing and control element (PCE) are also associated with USB.HOST 2.0,
USB.HOST 2.0 is connected with the upstream port of USB diverter modules.
The USB diverter modules include multiport transmitting-receiving control unit, and multiport transmitting-receiving control unit is connected to
Port to be swum, exchanges, change, cache, handle, USB1, USB3 and USB2, the camera of USB1 and image capture module connects,
The spare interface of USB3 and image capture module connects, and the USB interface of USB2 and communication module connects.
The communication module includes 4G communication chips, and 4G communication chips are connected to radio-frequency antenna, SIM card and USB and connect
Mouthful, radio-frequency antenna is connected with Cloud Server.
The Software for Design part of the present invention mainly includes the programming on Cloud Server and the journey of reading aid for blindmen front end
Sequence designs.Software for Design on Cloud Server be mainly the image of front-end collection is pre-processed, Text segmentation, finally utilize
The model trained completes the identification function of word, finally retransmits recognition result to front end main control chip.Blind person reads
The programming of device front end mainly text image collection, transmission and recognition result of the communication module to image including camera
The programs of functional module such as receiving, the configuration of report and keys interrupt of the voice broadcast module to recognition result are write.
Essence of the invention is that the character image that image capture module collects is identified.The key of Text region
Technology is to pre-process the image of collection, three links of Text segmentation and Text region, and a link is to image procossing
Effect is bad will to influence final recognition effect.
Picture pre-processes：
Due to needing that picture character is cut, so the gray value of the row of picture and line space (row and column pitch) should
The diversity ratio is larger, and noise is as far as possible small, is also exactly not have overlapping between row and row (row and arrange), whole word will not
There is inclination and distort.Therefore following 5 aspects are mainly included in terms of picture preconditioning technique：Gray processing, denoising, tilt school
Just with rim detection and Contrast enhanced.
Gray processing：
Image binaryzation, also referred to as gray processing, obvious black and white effect is exactly showed on image.As above us
Introduced, normal coloured image is a kind of spatial model with RGB3 passages, and for a pixel, we need simultaneously
Represent that black just can determine that a color to RGB triple channels assignment such as (0,0,0), and we are difficult to determine for a color
What RGB is respectively, and the digital operation of triple channel causes the treatment effeciency in pixel scale than relatively low, and if gray scale
Scheme this single channel image, we just readily can determine its pixel value for a kind of color.
By binary conversion treatment, we can be converted to picture in gray scale picture, so not only make it that treatment effeciency is higher,
And unique color pixel values are can determine so that the gray value differences of literal line and spaced rows are away from obvious.
After treatment, it is evident that it can be seen that word picture only has greyish black two kinds of colors, basically reached row with the ranks
It is obvious away from grey value difference before (row and column pitch).
Denoising：
Digital picture in reality is often made an uproar during transmission and digitlization by external environment condition and imaging device
Acoustic jamming etc. influences, referred to as noisy image or noise image.This process for reducing noise in digital picture is referred to as image
Make an uproar.Denoising can effectively improve the recognition effect of picture, and the method for denoising has a lot, and more commonly used mainly has mean filter
Device, adaptive wiener filter, median filter, Wavelet Denoising Method etc..
Denoising of the present invention is a kind of nonlinear smoothing technology using mainly median filtering method, median filtering method, and it will be every
The gray value of one pixel is arranged to the intermediate value of all pixels point gray value in the point neighborhood window.
Slant correction：
Take pictures or it is other acquisition picture during, can more or less cause the part of picture character to tilt, in order to
The degree of accuracy of the word by row segmentation is improved, entering line tilt correction plays the role of important, and how the key of picture Slant Rectify is
Detect the inclined direction of character and angle, and currently used slant correction algorithm mainly have it is following several：Conventional linear
The method of detection, the method based on projection, the method based on Hough transform and progress Fourier transformation are transformed into frequency domain progress
Analysis detection etc., by studying and testing this kind of characters such as word, numeral, letter of comprising only found for OCR identifications
Image is optimal using the method effect of Hough transform.
Hough takes a series of processing step such as highest tomographic image and then extraction image border by generating image pyramid
Then the angle of inclination for detecting image that can be more accurate and direction carry out rotation correction, test result shows this algorithm essence
Exactness highest.
Contrast enhanced：
For the average image of grey value profile, its visual effect is better than the image of other distributions, gray value
Uneven embodiment be background colour may gray value between a very big scope, and character color equally also has very big model
Enclose, for this phenomenon, we can be handled so as to highlight the contrast of character and background by the method for histogram equalization
Degree.
Effect when picture contrast is for picture segmentation has very big influence, if literal line is relative to line space
Contrast is stronger, histogram Wave crest and wave trough will significantly occurs in detection, on the contrary then be probably that a linear rise is right
The broken line declined afterwards, it is clear that appearance is smoother if pixel value has histogram Wave crest and wave trough, and we are just easy at these
Cut in place.So the purpose handled using Contrast enhanced seeks to strengthen the difference of the gray value of literal line and line space
And reduce the gray difference of its own.
The processing of contrast enhancing is carried out for improving recognition effect for the uneven picture of picture grey value profile
Effect is apparent.
Rim detection：
Rim detection, it is that brightness changes obvious point in reference numerals image, this is also image procossing and computer
Basic problem in vision, rim detection are almost essential in field of character recognition.
Image Edge-Detection is eliminated with identifying incoherent information, is known by significantly reducing data volume to save
The other time, and the important structure attribute of image is remained, improve word picture when reading aloud device identification importing and split by row
Efficiency.
Text segmentation：
By pretreatment above, we have obtained the binary picture of a contrast enhancing, below can to the figure
Piece split, and picture segmentation is broadly divided into following steps：Read gray-scale map and by its binaryzation, literal line segmentation,
Word column split, save as the small figures of single jpg.
Read gray-scale map and by its binaryzation：
Each pixel that picture is mainly stored in internal memory in the form of two-dimensional matrix, due to being gray-scale map so
What is now generated is single channel two-dimensional matrix.The colouring information of the in store picture of two-dimensional matrix, the span of color value is 0-
255, black correspond to 0, and then white correspond to 255, then the middle incremental color value that is uniformly scattered here and there.In Opencv
Mat matrix of the imread functions pictures reading into mat array gray-scale maps can be directly invoked.
Or most of numerical value can be evident that in 0-5, or in 250-255.0-5 numerical value this represent to carry on the back
The color value of scape, and the middle 250-255 being mingled with value, represent the color value of white font.Why so handling is
Prepared for the segmentation of subsequent pictures.
Text segmentation, by analysis, image at this stage is black matrix wrongly written or mispronounced character, and the pixel value of black is between 0-5, wrongly written or mispronounced character
Pixel between 250-255.In order to reach the purpose of separating character, it is necessary to which the pixel that pixel value is 0-5 is normalized into 0,
The pixel that pixel value is 250-255 is normalized to 255.The row of word can be so carried out to picture using the value of row pixel sum
Segmentation.If black background, one-row pixels and be necessarily 0 if there is white font, the pixel of a line and are naturally larger than 0.
Word can be split by row using this algorithm.
The often row character image arrived by row segmentation is preserved, is then gone out each Character segmentation using the method for row pixel sum
Come (if black background, a row pixel and necessarily be 0, if there is white font, one row pixel be naturally larger than 0).
May be to the Chinese character segmentation of tiled configuration into two parts using the method, therefore need to carry out word picture before it is split
Expansion process, column split then is being carried out, is being then converted into white gravoply, with black engraved characters effect.
Text region：
After above-mentioned figure pretreatment, Text segmentation operation is completed, next need that the word of segmentation is identified.
The acquisition of training data：
The good model of training one needs enough data.In the design, English character, Arabic numerals are completed
With the identification of Chinese characters in common use.Wherein, English character 52 kinds (A-Z, a-z), Arabic numerals 10 kinds (0-9), Chinese characters in common use 3500
Kind, but there is the fonts such as the Song typeface, black matrix in Chinese character, increase the difficulty of Chinese Character Recognition (only to the Song typeface, regular script, person in servitude in the design
Book, black matrix, children's circle, Chinese-language row pattern, the new Wei of Chinese, the easypro body of Chinese are trained).Training data is to write program using Python
The 40*40 of generation single character image, word corresponding to each font generate 5, collectively generate 142480 view data
As training set, then feasibility of 28496 images of generation as test set checking model at random.The corresponding generation of various fonts
Word is as shown in Figure 3.
The selection of training pattern：
Convolutional neural networks (Convolutional Neural Network, CNN) conduct in deep learning of the present invention
The training pattern of single character recognition.CNN (Fig. 4) is one kind of artificial neural network, but is more closely similar in network structure
The neutral net of organism.Compared to traditional artificial neural network, convolutional neural networks model on the one hand can be directly by original
Beginning image is inputted, and automatically extracts characteristics of image；Another aspect CNN models have more preferable generalization ability, even if working as image
Also recognition result will not be caused when deforming or noise be present significantly to influence；Another further aspect its to pass through local sensing wild
The complexity of network model is reduced with the method for shared weights, but it is higher than the accuracy of conventional model.
Training network model uses for reference the network of MNIST Handwritten Digit Recognitions, but Chinese character have it is increasingly complex and fine
Structure, therefore the adjustment of the network parameters such as parameter setting, network model configuration is carried out on this basis, make it more suitable for Chinese character
Training and identification.Training network model is as shown in Figure 5.
Neural unit in CNN between adjacent layer is not to connect entirely, but part connects, and the connection between neuron comes from
The partial nerve member of last layer.CNN models have used ReLU activation primitives used by this identifying system, and it belongs to unsaturation and swashed
Function living, due to when neutral net carries out error back propagation, gradient disappearance be present, i.e., every layer will be with activation primitive
First derivativeBe multiplied, when the network number of plies is more, gradient G will constantly decay until [0 ,+∞) disappear.
In terms of activation primitive, the design is using ReLu functions as activation primitive
Compared with traditional sigmoid, tanh function, ReLu functions largely can be with the effect of lift scheme.
To improve model generalization ability, add LRN layers (local acknowledgement normalizes layer).This layer of mimic biology nervous system
Lateral inhibition mechanism, the activity to local neuron creates competition mechanism so that it is relatively bigger to respond bigger value, performs one
Kind " lateral inhibition " operation, normalizing operation is done to the part for inputting number, smoothing processing is done in the output of current layer.The each input value of the layer
Will divided byN is local size size：Local_size, α are zoom factor, and β is exponential term, acquiescence
It is worth for 5, the calculation formula for doing smoothness constraint is:
Pooling layers in network use overlapping pool with max pattern, i.e., choose the certain area of input data in the calculation
The maximum in domain, records position of the maximum in each zonule, and during backpropagation, residual error is delivered into the maximum position
Put, other positions zero setting.Residual computations formula is：
If current layer k is convolutional layer, k+1 layers are next layer of sub-sampling layer, and the residual error of j-th of characteristic pattern of kth layer is public
Formula is：
Wherein, up (x) is to extend the size of+1 layer of kth as kth layer.
After N layer convolutional layer iteration, to prevent over-fitting, added in rear several layers of pooling layers and ReLU layers
Dropout layers, dropout layers it is random allow some nodes to export zero setting, also do not update weight, therefore some features can be avoided
Only just come into force under fixed combination, it is conscious to allow network to go to learn some universal general character, rather than some training samples
Some characteristics.
Last output layer selects SoftmaxWithLoss graders, as shown in Figure 6.SoftmaxWithLoss is really
Multinomial Logistic Loss Layer (cross entropy cost function) and Softmax Layer combination.Assuming that sample
Quantity has m, and each sample characteristics quantity is b, calculates this probability of m sample in n class, and calculation formula is：
By the real vector (a of k dimensions1,a2,a3,...,an) it is mapped as (b1,b2,b3,...,bn), then biAccording to size come
Carry out more classification tasks (weighting weight maximum one-dimensional).
Image capture module uses high-definition camera, and images to be recognized is sent into front end, is carried out by convolutional neural networks
Propagated forward, propagated forward general principle are as follows：
It is calculated as from input block to first hidden layer H1：Wherein k values travel through all input layers
Node, ZkIt is the weighted sum to all nodes of preceding layer, f () is nonlinear function, and rear layer is by that analogy.
Input information is handled from input layer through hidden layer, and is transmitted to output layer, one under the influence of the state of each layer of neuron
The state of layer neuron, for network output and the error of desired output, with the nonlinear interaction function of output layer.For input
Layer arrives the weight w of hidden layerijThen still using BP algorithm renewal weight.Input information is handled from input layer through hidden layer, specific processing
Mode is：Initial weight and thresholding are set, they are all set to less random number.Output layer is transmitted to after hidden layer is handled by defeated
Enter new input vectorWith corresponding output vectorPropagate.
In the propagated forward stage, data source arises from digital independent layer, by some process layers, reaches last layer (loss
Layer either characteristic layer), in this stage, the weights in network do not change, and networking path is a directed acyclic graph
(DAG), the node since most, by some process layers, in the absence of loop structure, therefore data flow can go ahead and push away
Enter until terminal.
Propagated forward process is studied with dataflow analysis method, i.e.,：From input data concentrate take a sample (X,
Y), wherein X is data, and Y is label.X is sent into network, successively calculated, obtains corresponding network processes output O, network performs
Calculating be formulated as：
O=Fn(...(F2(F1(XW1)W2)...)Wn)
Wherein, Fi, i-1,2, n represents nonlinear transformation, Wi, i=1,2, n represents each weights layer
Weights.O exports for network, (Y, O) can be used to assess network quality, preferable network meets Y==O.
In the training process of character data collection, information transmission will be inputted to output layer by first passing through propagated forward, to network
Output and error are modified, and last layer relatively obtains loss function with object function, calculation error updated value, adjusts hidden layer
Connection weight to output layer is：wjk=wjk-α(yk-tk)hj, α is learning rate.
Input layer is adjusted to the connection weight w of hidden layerij=wij-αWjxi, wherein
Because image feature information is more, data are numerous and diverse, therefore training process needs the training that iterates, until loss is received
Hold back, during training sample, keep uniformly input, can finally realize comparatively ideal discrimination.
In text categorization task, 3562 targets (Chinese characters in common use 3500, English upper and lower case letter 52 are classified as
Individual, Arabic numerals 10).Entered after training up, network structure model with iterations increase can reach 99% with
On accuracy (as shown in Figure 7), it is desirable to illustrate this network architecture.
The functional module that the present invention uses has central control module, image capture module, voice broadcast module, communication module
With USB diverter modules；
Central control module：The module will realize the number collected to whole reading aid for blindmen various pieces functional unit
Macro or mass analysis, and the scene reflected according to data are carried out according to information, corresponding control instruction is sent to each functional module.
According to the functional selection and demonstration to the design, it is desirable to which chip performance high power consumption is low, and suitable for doing hand-hold electronic equipments, communication is set
Standby or medical applications equipment, net book, learning machine, monitor video equipment and various man-machine interfaces, it can apply to high definition trip
Play, Wireless GPS, mobile video play, intelligent control, instrument and meter, navigation equipment, PDA device, remote monitoring, game
The exploitation such as exploitation.Core board can support HDMI voice harmony the core of the card piece voice synchronism outputs.Core board requirement at least meets 3
1G (1G dominant frequency, 1G internal memories, 1GFLASH) requirement, for running frequency up to 1GHZ, processor inside is 64/32 BITBUS network structure.
32/32KB level caches, 512KB L2 caches.Carry 3D figures accelerating engine (SGX540), 2D graphics accelerators, maximum branch
Hold 8192*8192 resolution ratio.Video coding supports MPEG-4/H.263/H. to reach 1080@30fps, decodes MPEG2/VC1/
Xvid videos reach [email protected] LINUX PDA operating systems are equipped with exclusively for core board in the present invention, make
Modules operational efficiency is higher, performance is more stable, real-time performance is stronger.Intelligent blind read core controls hardware block diagram
As shown in Figure 2.
Image capture module：After reader switch is started, corresponding button is selected to select corresponding recognition mode,
The central control module of reading aid for blindmen is sent to come the character image of shooting clear by 13,000,000 camera, then by pacifying
The main control chip of Linux system has been filled to be compressed to image, after the networking of 3g/4g modules, has been passed by SOCKET programs
To server, character image is identified server end, generates corresponding text, passes front-end collection system back again,
Reported finally by voice broadcast module, generate corresponding picture.
Camera uses MJPGE format compression images, and MJPEG refers to Motion JPEG, i.e. motion jpeg, according to 25 frames/
Second speed uses jpeg algorithm compressed video signal, completes the compression of dynamic video.Its picture format is that each frame is pressed
Contracting, can be usually reached 6：1 compression ratio, being all independent image just as each frame.Motion jpeg can produce high quality, complete
Screen, the video of full motion.The so picture pixels of collection out are higher and clear.
Socket using tcp host-host protocols, TCP (transmission control protocol) is a kind of connection-oriented network transmission
Agreement.Support multiple data stream operation, there is provided stream control and Wrong control, or even the rearrangement to out of order arrival message, therefore,
TCP transmission provides reliable data transport service.It is quick so make it that data transfer is stablized.
Voice broadcast module：Voice broadcast module includes pronounciation processing chip (SYN6288) and other peripheral circuits, the language
The major function of sound processing module is exactly to realize the man-machine interaction of user and various functions intermodule, when user uses blind person
During reader, after the text message collected is identified, obtained text information is sent to voice broadcast module, the mould
The data that block then needs to obtain are handled, and are reported by outside speaker or ear speaker device, inform user's word
Interior lane, complete the reading of the printing words such as books.The peripheral circuit schematic diagram of its pronounciation processing chip is as indicated at 3.
Communication module：Communication module is 4G communication modules, the realization of network communicating function, make use of 4G modules without line width
Band network access facility.Fourth generation mobile phone mobile communication standard, refer to fourth generation mobile communication technology, foreign language abbreviation：
4G.The technology includes two kinds of standards of TD-LTE and FDD-LTE, is to integrate 3G and WLAN, and can quickly transmit data, height
Quality, audio, video and image etc..4G can be downloaded with more than 100Mbps speed, than current home broadband AD SL (4
Million) fast 25 times, and disclosure satisfy that requirement of nearly all user for wireless service.In addition, 4G can be in DSL and cable television
The place deployment that modem is not covered with, then expands to whole distract again.
4G modules are connected by USB interface with central control module, and USB main lines bus signals are between each module and system
Communication interface.ME3760_V2 modules are loaded according to the ECM mouths carried the driving of Linux3.2.0 kernels, used
The USB port of ME3760_V2 modules is attached, and under linux, the ECM mouths of ME3760_V2 modules are mapped to 5 interfaces：
ECM, AT, Modem, Log, wherein " " parts of ECM mouths is fallen within, to prevent ECM functions to be capped, carrying out USB strings
It should be filtered during mouth initialization, and load the PPP drivings of linux kernel, remaining interface initialization is set for USB serial ports
It is standby, finally with PPP instrument dial-up connection 4G networks.
The present invention is realized using 4G communication modules carries out two-way communication between reading aid for blindmen front end and Cloud Server.When
The text image that camera gathers is sent to Cloud Server to complete Text region, second, Cloud Server sends out the result of identification
The main control chip of reading aid for blindmen front end is delivered to, recognition result is informed by user by voice broadcast module in real time, completes to read
Read function.Intelligent blind reader 4G communication module application circuit schematic diagrams are as shown in Figure 4.
USB diverter modules：There are two parts through what USB interface was connected with central control module, first, camera, connects with USB
Mouth one is connected；Second, 4G communication modules, are connected with USB interface two.Two parts are respectively through USB diverter modules and central control module
Connection, carry out information exchange transmission.
FE1.1s is a highly integrated, high quality, high-performance, low energy consumption, at the same still USB2.0 at a high speed 4 port line concentrations it is low
Cost solution.It using it is single exchange converter (Single Transaction Translator) (STT) structure so as to
Obtain more benefits.Six rather than two aperiodic translation caches are to reduce potential Communication Jamming.Whole design is based on
State machine controls, to reduce the time delay of response；Microcontroller is not used in this chip.To ensure high quality, entirely
Chip makes all logic elements by testing scan chain (Test Scan Chain)-under at a high speed (480MHz) pattern
Fully tested before shipment.Particularly Built In Self Test (Build-In-Self-Test) type order be use all height
Speed, at full speed, and low-speed mode AFE(analog front end) port (AFE) is also such in the packaging and testing stage.The realization of low energy consumption is
By using 0.18 μm of technology and integrated power supply/clock control mechanism.Most of pin does not need timing, unless by with
Arrive.USB diverter module fundamental diagrams are as shown in Figure 5.
The present invention operation principle be：Image capture module gathers text image information, send USB diverter modules, USB shuntings
Text image information is uploaded to Cloud Server by module by communication module, and Cloud Server is using digital figure treatment technology to adopting
The image information of collection is pre-processed, the processing of Text segmentation, Text region, then the model by having trained completes Text region work(
Can, and recognition result is sent to USB diverter modules with text formatting, USB diverter modules are then forwarded to central control module, in
The recognition result of text formatting is changed into voice messaging, language by centre control module by the Audio Processing Unit of voice broadcast module
Message breath send bluetooth unit, by the earphone or loudspeaker of bluetooth unit, by voice broadcast to reader.
Claims (9)
- A kind of 1. intelligent blind reading method, it is characterised in that：The image information of text is gathered by image capture module, by The image information of text is uploaded to Cloud Server by USB diverter modules by communication module；The image information of Cloud Server collection Pre-processed, Text segmentation, Text region, and recognition result is sent to USB diverter modules with text formatting；USB is shunted Module is retransmited to central control module, and central control module is by the Audio Processing Unit of voice broadcast module by text formatting Recognition result change into voice messaging, bluetooth unit is sent, by the earphone or loudspeaker of bluetooth unit, by voice broadcast to being read Reader.
- 2. intelligent blind reading method according to claim 1, it is characterised in that：The Cloud Server is by image point Analysis system completes picture pretreatment, Text segmentation and Text region.
- 3. intelligent blind reading method according to claim 2, it is characterised in that：Described Text region, it is using text Word train model is completed, the training pattern using the convolutional neural networks in deep learning as single character recognition.
- 4. according to the intelligent blind reader system of the intelligent blind reading method of claim 1,2 or 3, its feature exists In：Including central control module, central control module is connected to voice broadcast module and USB diverter modules, USB divergent dies Block is connected by communication module with Cloud Server, and USB diverter modules are also associated with image capture module.
- 5. intelligent blind reader system according to claim 4, it is characterised in that：The voice broadcast module includes language Sound processing unit, Audio Processing Unit are connected with bluetooth unit.
- 6. intelligent blind reader system according to claim 5, it is characterised in that：The phonetic synthesis unit includes string Mouth transmitting-receiving, phonetic synthesis and voice output circuit, the bluetooth unit include power amplifier and loudspeaker；The input of serial ports transmitting-receiving End is connected with the UART3 of central control module, and the output end of serial ports transmitting-receiving passes through phonetic synthesis and voice output connection, language The power amplifier that the output of sound outlet line passes sequentially through bluetooth unit connects with loudspeaker.
- 7. intelligent blind reader system according to claim 4, it is characterised in that：The central control module includes master Chip is controlled, main control chip is connected with UART3 and UART2 respectively by caching, processing and control element (PCE), and UART2 is connected with reserved string Mouthful, caching, processing and control element (PCE) are also associated with USB.HOST 2.0, USB.HOST 2.0 and USB diverter modules upstream port Connection.
- 8. intelligent blind reader system according to claim 4, it is characterised in that：The USB diverter modules include more Port transmitting-receiving control unit, multiport transmitting-receiving control unit are connected to upstream port, exchange, change, caching, processing, The spare interface of the camera connection of USB1, USB3 and USB2, USB1 and image capture module, USB3 and image capture module connects Connect, the USB interface connection of USB2 and communication module.
- 9. intelligent blind reader system according to claim 4, it is characterised in that：The communication module communicates including 4G Chip, 4G communication chips are connected to radio-frequency antenna, SIM card and USB interface, and radio-frequency antenna is connected with Cloud Server.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710739516.1A CN107346629A (en) | 2017-08-22 | 2017-08-22 | A kind of intelligent blind reading method and intelligent blind reader system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710739516.1A CN107346629A (en) | 2017-08-22 | 2017-08-22 | A kind of intelligent blind reading method and intelligent blind reader system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107346629A true CN107346629A (en) | 2017-11-14 |
Family
ID=60257562
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710739516.1A Pending CN107346629A (en) | 2017-08-22 | 2017-08-22 | A kind of intelligent blind reading method and intelligent blind reader system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107346629A (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108304814A (en) * | 2018-02-08 | 2018-07-20 | 海南云江科技有限公司 | A kind of construction method and computing device of literal type detection model |
| CN108363522A (en) * | 2018-04-24 | 2018-08-03 | 石家庄科达文教用品有限公司 | Synchronous writing system and its method |
| CN108551575A (en) * | 2018-06-25 | 2018-09-18 | 汉字造梦（武汉）文化传播有限公司 | A kind of glyphomancy interactive experience system and method |
| CN108764261A (en) * | 2018-05-31 | 2018-11-06 | 努比亚技术有限公司 | A kind of image processing method, mobile terminal and storage medium |
| CN109062404A (en) * | 2018-07-20 | 2018-12-21 | 东北大学 | A kind of interactive system and method applied to intelligent children's early learning machine |
| CN109272043A (en) * | 2018-09-21 | 2019-01-25 | 北京京东金融科技控股有限公司 | Training data generation method, system and electronic equipment for optical character identification |
| CN110222684A (en) * | 2019-04-19 | 2019-09-10 | 黑龙江大学 | A kind of blind person's " reading " system |
| CN110287830A (en) * | 2019-06-11 | 2019-09-27 | 广州市小篆科技有限公司 | Intelligence wearing terminal, cloud server and data processing method |
| CN110334712A (en) * | 2019-06-11 | 2019-10-15 | 广州市小篆科技有限公司 | Intelligence wearing terminal, cloud server and data processing method |
| CN110705556A (en) * | 2019-09-18 | 2020-01-17 | 佛山职业技术学院 | Braille conversion instrument, conversion method and storage medium |
| CN111027345A (en) * | 2018-10-09 | 2020-04-17 | 北京金山办公软件股份有限公司 | Font identification method and apparatus |
| CN111126341A (en) * | 2019-12-31 | 2020-05-08 | 中山大学新华学院 | Service system for blind people |
| CN111814800A (en) * | 2020-07-24 | 2020-10-23 | 广州广杰网络科技有限公司 | Aged book and newspaper reader based on 5G + AIoT technology and use method thereof |
| CN114627730A (en) * | 2022-03-31 | 2022-06-14 | 北京科技大学 | Braille electronic book |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103077625A (en) * | 2013-01-30 | 2013-05-01 | 中国盲文出版社 | Blind electronic reader and blind assistance reading method |
| CN104966084A (en) * | 2015-07-07 | 2015-10-07 | 北京奥美达科技有限公司 | OCR (Optical Character Recognition) and TTS (Text To Speech) based low-vision reading visual aid system |
| CN106341549A (en) * | 2016-10-14 | 2017-01-18 | 努比亚技术有限公司 | Mobile terminal audio reading apparatus and method |
| CN106406445A (en) * | 2016-09-09 | 2017-02-15 | 华南理工大学 | Intelligent glasses-based Chinese text reading system assisting in visual impairment |
| CN106408038A (en) * | 2016-09-09 | 2017-02-15 | 华南理工大学 | Rotary Chinese character identifying method based on convolution neural network model |
| CN106389078A (en) * | 2016-11-24 | 2017-02-15 | 贵州大学 | Intelligent blind guiding glass system and blind guiding method thereof |
-
2017
- 2017-08-22 CN CN201710739516.1A patent/CN107346629A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103077625A (en) * | 2013-01-30 | 2013-05-01 | 中国盲文出版社 | Blind electronic reader and blind assistance reading method |
| CN104966084A (en) * | 2015-07-07 | 2015-10-07 | 北京奥美达科技有限公司 | OCR (Optical Character Recognition) and TTS (Text To Speech) based low-vision reading visual aid system |
| CN106406445A (en) * | 2016-09-09 | 2017-02-15 | 华南理工大学 | Intelligent glasses-based Chinese text reading system assisting in visual impairment |
| CN106408038A (en) * | 2016-09-09 | 2017-02-15 | 华南理工大学 | Rotary Chinese character identifying method based on convolution neural network model |
| CN106341549A (en) * | 2016-10-14 | 2017-01-18 | 努比亚技术有限公司 | Mobile terminal audio reading apparatus and method |
| CN106389078A (en) * | 2016-11-24 | 2017-02-15 | 贵州大学 | Intelligent blind guiding glass system and blind guiding method thereof |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108304814B (en) * | 2018-02-08 | 2020-07-14 | 海南云江科技有限公司 | Method for constructing character type detection model and computing equipment |
| CN108304814A (en) * | 2018-02-08 | 2018-07-20 | 海南云江科技有限公司 | A kind of construction method and computing device of literal type detection model |
| CN108363522A (en) * | 2018-04-24 | 2018-08-03 | 石家庄科达文教用品有限公司 | Synchronous writing system and its method |
| CN108764261A (en) * | 2018-05-31 | 2018-11-06 | 努比亚技术有限公司 | A kind of image processing method, mobile terminal and storage medium |
| CN108551575A (en) * | 2018-06-25 | 2018-09-18 | 汉字造梦（武汉）文化传播有限公司 | A kind of glyphomancy interactive experience system and method |
| CN109062404B (en) * | 2018-07-20 | 2020-03-24 | 东北大学 | Interaction system and method applied to intelligent early education machine for children |
| CN109062404A (en) * | 2018-07-20 | 2018-12-21 | 东北大学 | A kind of interactive system and method applied to intelligent children's early learning machine |
| CN109272043A (en) * | 2018-09-21 | 2019-01-25 | 北京京东金融科技控股有限公司 | Training data generation method, system and electronic equipment for optical character identification |
| CN111027345A (en) * | 2018-10-09 | 2020-04-17 | 北京金山办公软件股份有限公司 | Font identification method and apparatus |
| CN110222684A (en) * | 2019-04-19 | 2019-09-10 | 黑龙江大学 | A kind of blind person's " reading " system |
| CN110334712A (en) * | 2019-06-11 | 2019-10-15 | 广州市小篆科技有限公司 | Intelligence wearing terminal, cloud server and data processing method |
| CN110287830A (en) * | 2019-06-11 | 2019-09-27 | 广州市小篆科技有限公司 | Intelligence wearing terminal, cloud server and data processing method |
| CN110705556A (en) * | 2019-09-18 | 2020-01-17 | 佛山职业技术学院 | Braille conversion instrument, conversion method and storage medium |
| CN111126341A (en) * | 2019-12-31 | 2020-05-08 | 中山大学新华学院 | Service system for blind people |
| CN111814800A (en) * | 2020-07-24 | 2020-10-23 | 广州广杰网络科技有限公司 | Aged book and newspaper reader based on 5G + AIoT technology and use method thereof |
| CN114627730A (en) * | 2022-03-31 | 2022-06-14 | 北京科技大学 | Braille electronic book |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107346629A (en) | A kind of intelligent blind reading method and intelligent blind reader system | |
| CN110276316B (en) | Human body key point detection method based on deep learning | |
| CN106682569A (en) | Fast traffic signboard recognition method based on convolution neural network | |
| CN108875708A (en) | Behavior analysis method, device, equipment, system and storage medium based on video | |
| CN108133188A (en) | A kind of Activity recognition method based on motion history image and convolutional neural networks | |
| CN106778506A (en) | A kind of expression recognition method for merging depth image and multi-channel feature | |
| CN108985217A (en) | A kind of traffic sign recognition method and system based on deep space network | |
| CN108806792A (en) | Deep learning facial diagnosis system | |
| CN106951867A (en) | Face identification method, device, system and equipment based on convolutional neural networks | |
| CN107491726A (en) | A kind of real-time expression recognition method based on multi-channel parallel convolutional neural networks | |
| CN108052884A (en) | A kind of gesture identification method based on improvement residual error neutral net | |
| CN108764195A (en) | Handwriting model training method, hand-written character recognizing method, device, equipment and medium | |
| CN109543632A (en) | A kind of deep layer network pedestrian detection method based on the guidance of shallow-layer Fusion Features | |
| CN110458060A (en) | A kind of vehicle image optimization method and system based on confrontation study | |
| Teixeira et al. | Address-event imagers for sensor networks: evaluation and modeling | |
| CN109817276A (en) | A kind of secondary protein structure prediction method based on deep neural network | |
| CN107704859A (en) | A kind of character recognition method based on deep learning training framework | |
| CN104778466B (en) | A kind of image attention method for detecting area for combining a variety of context cues | |
| CN110490239A (en) | Training method, the quality classification method, device and equipment of image quality control network | |
| CN108009548A (en) | A kind of Intelligent road sign recognition methods and system | |
| CN112419326B (en) | Image segmentation data processing method, device, equipment and storage medium | |
| CN110008961A (en) | Text real-time identification method, device, computer equipment and storage medium | |
| CN109086653A (en) | Handwriting model training method, hand-written character recognizing method, device, equipment and medium | |
| CN110046544A (en) | Digital gesture identification method based on convolutional neural networks | |
| CN110968235B (en) | Signal processing device and related product |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171114 |