CN110309846A - A kind of image-recognizing method based on deep space domain network - Google Patents
A kind of image-recognizing method based on deep space domain network Download PDFInfo
- Publication number
- CN110309846A CN110309846A CN201910342636.7A CN201910342636A CN110309846A CN 110309846 A CN110309846 A CN 110309846A CN 201910342636 A CN201910342636 A CN 201910342636A CN 110309846 A CN110309846 A CN 110309846A
- Authority
- CN
- China
- Prior art keywords
- image
- fft
- present
- training
- test
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image-recognizing methods based on deep space domain network.Traditional machine learning method different from the past is to have good recognition effect to the image of fixed scene, images to be recognized is switched to spatial domain using FFT technique first by the present invention, according to perception theory, deep thoughts can more accurately understand problem, the present invention invents three layers of FFT network by this thinking, first time is extracted space domain characteristic as the input of the second layer and is carrying out FFT feature extraction by the present invention, obtain input of the feature as third layer, carry out FFT feature extraction again later, the FFT characteristic image of the characteristics of image extracted and mark is carried out match cognization with softmax classifier by the present invention later.
Description
Technical field
The present invention relates to computer vision, mathematical image processing and models to identify field, more particularly to a kind of based on depth
The image-recognizing method of spatial domain network.
Background technique
Biometrics is model identification and one important branch of computer vision.Image recognition technology is bio-measurement
It learns one of most crucial technology and it has been widely used in public security investigation, space flight exploration and access control system.
However, image recognition technology still suffers from some challenges, by taking face recognition technology as an example, these technologies are in variation lamp
Light, when blocking with expression shape change, recognition effect can all reduce.Therefore, how preferably to indicate that image is known as the hot spot studied.
Whole representation is to be proposed for image recognition earliest, it extracts characteristic point mainly to represent entirety, such as
One width facial image replaces whole face with eyes, eyebrow, nose and mouth information, and this method loses relationship between specified point, very
It is abandoned fastly by everybody.Dimension reduction method is generalized quickly later, and Principal Component Analysis (PCA) is most typical dimension reduction method, it
Image is indicated using vector, the average value and covariance of every group of vector is then found out, is indicated using covariance feature original
Image, this method obtain very good effect in image recognition.But this method relationship between vector suddenly, loses crucial information.
It ties up principal component analytical method later in order to solve the method 2 to be suggested, when this method uses a matrix to indicate image, then finds out
The covariance of matrix, using the feature of covariance as the feature of image, the dimension of image is effectively reduced in this, while in image recognition
Upper acquisition good performance.But this method has high complexity.In order to reduce complexity, Sparse methods are suggested.Especially
Good performance is obtained on solving Small Sample Database collection, specifically: given test is indicated using all training samples
Sample, and sparse coefficient is found out, classified using sparse coefficient and the reconstruct of every class training sample with all test samples, it should
Method is referred to as 1 norm Sparse methods, and this method, which solves sparse solution using 1 norm, has high complexity.It is asked to solve this
Topic, 2 norm Sparse methods are suggested, and no matter all well received in the performance or efficiency of image recognition this method is.But it should
Method can only handle the image of special scenes, such as: medical image.This limits its application significantly.
The present invention proposes a kind of image-recognizing method based on deep space domain network and device for image recognition, the party
Method is obtained the feature of robust by multilayer spatial domain learning characteristic and carries out Classification of Colleges using softmax.This method energy
Effectively handle the general image recognition tasks including specific image identification.
Summary of the invention
The present invention proposes a kind of image-recognizing method based on deep space domain, and this method can effectively handle traditional machine
Device learning method can also identify the image under the specific environments such as medicine, remote sensing to the deficiency of image recognition under complex environment.
Solution proposed by the present invention is as follows:
1. overall flow figure of the invention, referring to a kind of process of the image-recognizing method based on deep space domain network of attached drawing 1
Figure.
2. being specially training set by the half of every a kind of image, remaining image is by image classification training set and test set
Test set.The data set that the present invention uses is YALE, GT, AR and COIL.
3. training image is rotated by 90 ° by the present invention, 180 °, 270 ° of progress image expansions, width training image is just one by one in this way
Extension becomes the image of 4 width training.
4. training image and test image pass through FFT respectively extracts spatial domain.Spatial domain can more preferably reflect the distribution of image.
The network that the present invention designs has three layers, and every layer is made of FFT.
5. training image passes through input of the FFT image as the first layer network, obtained feature is denoted as F1;
6 F1 pass through the second layer network being made of FFT, and the feature extracted is denoted as F2;
7.F2 passes through the third layer network being made of FFT, and the feature extracted is denoted as F3;
8.4,5,6,7 the specific implementation formula of fft algorithm involved in is as follows:
(1)
Wherein,It is to represent a width original image, the size of diagram picture is,It is that piece image is converted to frequency
Character representation behind domain,It is image slices vegetarian refreshments coordinate,The pixel of frequency domain image.
9.F3 and classification pass through Softmax and objective function is trained model;Wherein Softmax function is accomplished by
(2)
In formula (2),, sample is input, and output is usedIt is a as a result, which probability of outcome is big, which it just belongs to
Class.
10. function is least square error MSE;Here repetitive exercise number is 50 times;Shown in the realization of MSE such as formula (3):
(3)
In formula (3), if N represents data amount check, r representative group number,RepresentThe sample variance of group.If test specimens
Which group difference of this jade is smaller, which class is just belonged to when it is trained to.
11. testing the performance that the present invention proposes model with test data after the completion of training pattern in 7;
12. the present invention can be used for the image detections tasks such as medical image is diagnosed a disease, Face datection and license plate number detect.
It illustrates
The present invention proposes the image-recognizing method of a deep space domain network, and this method can handle the figure under complex environment
As identification, method proposed by the invention is shown in order to apparent, is illustrated.
1. every a kind of half of database is divided into training set and test set by the present invention, such as COIL100 data
Collection, one shares 100 classes, and every class has 355 images, and preceding 177 images are to train, and latter 178
Image is test image.Training set has 17700 images altogether, and test set has 17800 images.
2. every training image is carried out 90 ° by the present invention, 180 °, 270 ° of image rotations enhance training set, referring to attached drawing
Primitive beginning image graph, is rotated by 90 ° image in 2, rotates 180 ° of images, rotates 270 ° of images.
3. we convert space domain characteristic image using FFT method for training image and test image;
Referring to the original training image 1 of attached drawing 3, original training image 2, original test image 1, original test image 2;
FFT training image 1 is converted to referring to attached drawing 4, converts FFT test image 2, converts FFT test image 1;
F1 image is obtained by first layer referring to 5 training image 1 of attached drawing, tests F1 image 2, test image F1 image 1;
It is inputted referring to 6 F1 image of attached drawing as the second layer, obtains F2 image 1 by the second layer and test the test F2 image of F2 image 2
1;
Input referring to 7 F2 image of attached drawing as third layer obtains F3 image 1 by third layer and tests 2 F3 image 2 of F3 image
Test F3 image 1.
4. it is used to undated parameter by MSE, present invention understands that training is matched with given classification, and which
A which parameter that just updated determines model.
5. the present invention waits test 1 to be into model with test 2 with model is trained in 4 come verification the verifying results,
Network is counted in carrying out 2 probability by softmax, it is assumed that this probability is 0.2 and 0.8, and probability is exactly this test chart greatly
The classification of picture exactly tests 2 and belongs to the second class.Obtain figure in 3 from the graph it is also known that test 2 classification.
Detailed description of the invention
Specific embodiments of the present invention will be further explained with reference to the accompanying drawing:
A kind of flow chart of the image-recognizing method based on deep space domain network of Fig. 1;
Fig. 2 is original image figure, is rotated by 90 ° image, rotates 180 ° of images, rotates 270 ° of images;
Fig. 3 is original training image 1, original training image 2, original test image 1, original test image 2;
Fig. 4 is converted into FFT training image 1, converts FFT test image 2, converts FFT test image 1;
Fig. 5 is that the process first layer of training image 1 obtains F1 image, tests F1 image 2, test image F1 image 1;
Fig. 6 is that F1 image is inputted as the second layer, obtains F2 image 1 by the second layer and tests the test F2 image 1 of F2 image 2;
Fig. 7 is input of the F2 image as third layer, obtains F3 image 1 by third layer and tests the survey of 2 F3 image 2 of F3 image
Try F3 image 1.
Claims (11)
1. a kind of image-recognizing method based on deep space domain network, diagnose a disease and general image for handling medical image
Identification.
2. being specially training set by the half of every a kind of image, remaining image is by image classification training set and test set
Test set, the data set that the present invention uses is YALE, GT, AR and COIL.
3. training image is rotated by 90 ° by the present invention, and 180 °, 270 ° of progress image expansions.
4. training image and test image pass through FFT respectively extracts spatial domain, spatial domain can more preferably reflect the distribution of image,
The network that the present invention designs has three layers, and every layer is made of FFT.
5. training image passes through input of the FFT image as the first layer network, obtained feature is denoted as F1.
6. F1 passes through the second layer network being made of FFT, the feature extracted is denoted as F2.
7. F2 passes through the third layer network being made of FFT, the feature extracted is denoted as F3.
8. F3 and classification are trained model by Softmax and objective function.
9. objective function is least square error MSE;Here repetitive exercise number is 50 times.
10. testing the performance that the present invention proposes model with test data after the completion of training pattern in 7.
11. the present invention can be used for the image detections tasks such as medical image is diagnosed a disease, Face datection and license plate number detect.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910342636.7A CN110309846A (en) | 2019-04-26 | 2019-04-26 | A kind of image-recognizing method based on deep space domain network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910342636.7A CN110309846A (en) | 2019-04-26 | 2019-04-26 | A kind of image-recognizing method based on deep space domain network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110309846A true CN110309846A (en) | 2019-10-08 |
Family
ID=68074565
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910342636.7A Pending CN110309846A (en) | 2019-04-26 | 2019-04-26 | A kind of image-recognizing method based on deep space domain network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110309846A (en) |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101447021A (en) * | 2008-12-30 | 2009-06-03 | 爱德威软件开发(上海)有限公司 | Face fast recognition system and recognition method thereof |
CN105247539A (en) * | 2013-04-08 | 2016-01-13 | 考吉森公司 | Method for gaze tracking |
US20170357879A1 (en) * | 2017-08-01 | 2017-12-14 | Retina-Ai Llc | Systems and methods using weighted-ensemble supervised-learning for automatic detection of ophthalmic disease from images |
CN107909095A (en) * | 2017-11-07 | 2018-04-13 | 江苏大学 | A kind of image-recognizing method based on deep learning |
CN108351984A (en) * | 2015-11-05 | 2018-07-31 | 微软技术许可有限责任公司 | The depth convolutional neural networks of hardware-efficient |
CN109190544A (en) * | 2018-08-27 | 2019-01-11 | 华中科技大学 | A kind of human body personal identification method based on sequence depth image |
CN109522905A (en) * | 2018-10-18 | 2019-03-26 | 湖南大学 | A kind of hot spot detecting method based on FFT feature extraction |
-
2019
- 2019-04-26 CN CN201910342636.7A patent/CN110309846A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101447021A (en) * | 2008-12-30 | 2009-06-03 | 爱德威软件开发(上海)有限公司 | Face fast recognition system and recognition method thereof |
CN105247539A (en) * | 2013-04-08 | 2016-01-13 | 考吉森公司 | Method for gaze tracking |
CN108351984A (en) * | 2015-11-05 | 2018-07-31 | 微软技术许可有限责任公司 | The depth convolutional neural networks of hardware-efficient |
US20170357879A1 (en) * | 2017-08-01 | 2017-12-14 | Retina-Ai Llc | Systems and methods using weighted-ensemble supervised-learning for automatic detection of ophthalmic disease from images |
CN107909095A (en) * | 2017-11-07 | 2018-04-13 | 江苏大学 | A kind of image-recognizing method based on deep learning |
CN109190544A (en) * | 2018-08-27 | 2019-01-11 | 华中科技大学 | A kind of human body personal identification method based on sequence depth image |
CN109522905A (en) * | 2018-10-18 | 2019-03-26 | 湖南大学 | A kind of hot spot detecting method based on FFT feature extraction |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110348319B (en) | Face anti-counterfeiting method based on face depth information and edge image fusion | |
Patel et al. | Dictionary-based face recognition under variable lighting and pose | |
CN105138998B (en) | Pedestrian based on the adaptive sub-space learning algorithm in visual angle recognition methods and system again | |
CN105117708A (en) | Facial expression recognition method and apparatus | |
Chandran et al. | Missing child identification system using deep learning and multiclass SVM | |
CN110738071A (en) | face algorithm model training method based on deep learning and transfer learning | |
CN103593648B (en) | Face recognition method for open environment | |
Pratama et al. | Face recognition for presence system by using residual networks-50 architecture | |
Haji et al. | Real time face recognition system (RTFRS) | |
CN106980825A (en) | A kind of face posture sorting technique based on normalization pixel difference feature | |
Tahir et al. | Novel face detection algorithm with a mask on neural network training | |
CN116612335A (en) | Few-sample fine-granularity image classification method based on contrast learning | |
CN110502961A (en) | A kind of facial image detection method and device | |
Kadhim et al. | A multimodal biometric database and case study for face recognition based deep learning | |
Begum et al. | A novel approach for multimodal facial expression recognition using deep learning techniques | |
CN103984918B (en) | Human face image recognition method based on intra-class and inter-class variation | |
CN110309846A (en) | A kind of image-recognizing method based on deep space domain network | |
Janarthanan et al. | An Efficient Face Detection and Recognition System Using RVJA and SCNN | |
Sreelekshmi et al. | Deep forgery detect: enhancing social media security through deep learning-based forgery detection | |
Prabakar et al. | Hybrid deep learning model for copy move image forgery detection | |
Rasheed et al. | A Novel Method for Signature Verification Using Deep Learning | |
Charishma et al. | Smart Attendance System with and Without Mask using Face Recognition | |
CN111429156A (en) | Artificial intelligence recognition system for mobile phone and application thereof | |
Shylaja et al. | Illumination Invariant Novel Approaches for Face Recognition | |
CN111353443A (en) | Cross-view kernel collaborative representation-based pedestrian re-identification method |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20191008 |