CN111881312B - Image data set classification and division method - Google Patents
Image data set classification and division method Download PDFInfo
- Publication number
- CN111881312B CN111881312B CN202010722578.3A CN202010722578A CN111881312B CN 111881312 B CN111881312 B CN 111881312B CN 202010722578 A CN202010722578 A CN 202010722578A CN 111881312 B CN111881312 B CN 111881312B
- Authority
- CN
- China
- Prior art keywords
- image data
- hash
- vector
- data set
- feature
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/55—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Abstract
The invention discloses a method for classifying and dividing an image data set, which comprises the following steps: building a pre-training model, generating a projection matrix group R, zooming the image data of the picture, and importing the image data into the pre-training model; performing feature extraction on picture image data through a pre-training model to generate a feature vector x with a certain dimension; performing L2 regularization on each generated feature vector x, and scaling the feature vectors into unit vectors in a high-dimensional spherical space; for each projection matrix Ri, calculating the matrix to obtain a result vector, and taking an index corresponding to the maximum value in the result vector as a hash value hi of the characteristic vector; and combining all the calculated matrix hash values hi to obtain a group of hash values, taking the hash values as the hashes of the image characteristic quantities, and dividing the images with the same hash values into one class. By the scheme, the image features can be effectively extracted, the image data set is accurately divided, and the probability of Hash collision is reduced.
Description
Technical Field
The invention relates to the field of deep learning image data set classification processing, in particular to an image data set classification dividing method.
Background
With the development of multimedia technology, image classification has become the key point of research in the field of computer vision, image classification is to divide images into different preset categories according to certain attributes of the images, how to effectively express the images is the key point for improving the accuracy of image classification, and the problem of selecting and extracting features is the difficult problem existing in image classification at present. With the rapid development of mobile internet, human society has entered the big data era. Although some features of the images can be extracted through traditional feature learning such as SIFT and HOG, and a good effect is achieved in image classification, the artificial feature design method has certain defects. The existing image classification technology is not accurate enough in dividing an image data set, and when a Hash algorithm is used for processing image data, the probability of Hash collision cannot be effectively reduced.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide the image data set classification and division method, which can effectively extract image features, accurately divide the image data set and reduce the probability of Hash collision.
The purpose of the invention is realized by the following technical scheme:
a method for classifying and dividing an image data set comprises the following steps:
s1, building a pre-training model, randomly generating a projection matrix group R, zooming the image data of the pictures without the classification marks, and importing the image data into the pre-training model;
s2, performing feature extraction on the imported image data without the classified pictures through a pre-training model to generate a feature vector x with a certain dimensionality;
s3, performing L2 regularization on each generated feature vector x by using a regularization formula, and scaling the feature vectors into unit vectors in a high-dimensional spherical space;
s4, for each projection matrix Ri, calculating the matrix by using a Hash function calculation formula to obtain a result vector, and taking an index corresponding to the maximum value in the result vector as a Hash value hi of the feature vector;
s5, all the hash values hi calculated by the projection matrix group R in step S4 are combined to obtain a group of hash values and the group of hash values is used as a hash of the image feature quantity, and at this time, the images having the same hash value are classified into one type.
Specifically, the size of the projection matrix set in step S1 is 2048 × B, where B is the set hash bucket size.
Specifically, the pre-training model in step S1 is a ResNet50 model.
Specifically, the size of the scaled picture image data in step S1 is 224 × 224.
Specifically, the regularization formula in step S3 is:
wherein x isiRepresenting the ith feature in the x vector.
Specifically, the hash function calculation formula in step S4 is as follows: hi ═ argmax (x × R)i)。
Specifically, the hash of the image feature vector in step S5 is represented as: hash ═ h0,h1,h2,...]。
The invention has the beneficial effects that: by the scheme, the image features can be effectively extracted, the image data set is accurately divided, and the probability of Hash collision is reduced.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
FIG. 2 is a diagram of a pre-trained model architecture of the present invention.
FIG. 3 is a schematic diagram of the image feature vector classification process according to the present invention.
Detailed Description
In order to more clearly understand the technical features, objects, and effects of the present invention, embodiments of the present invention will now be described with reference to the accompanying drawings.
The image data set is classified and divided, and the classification result is that each unit belongs to a certain class, but the classification result is not allowed to belong to the class or belong to another class or is omitted. Such as classifying images as people, cats, dogs, tables, etc., the images in each class belong strictly to the class and not to the other classes, and the sum of the number of images in all classes is equal to the sum of the overall images.
In this embodiment, the present invention mainly provides a method for classifying and dividing an image data set for an image without a classification mark, as shown in fig. 1, the method includes the following steps:
And 2, performing feature extraction on the imported image data of the non-classified pictures through a pre-training model to generate a 2048-dimensional feature vector x.
And 3, performing L2 regularization on each generated feature vector x by using a regularization formula, and scaling the feature vectors into unit vectors in a high-dimensional spherical space. Wherein the regularization formula is shown as follows:
wherein x isiRepresenting the ith feature in the x vector.
And 4, calculating a formula for each projection matrix Ri by using a hash function: hi ═ argmax (x × R)i) And calculating the matrix to obtain a result vector, and taking an index corresponding to the maximum numerical value in the result vector as the hash value hi of the characteristic vector.
Step 5, combining all the hash values hi calculated by the projection matrix set R in step S4 to obtain a set of hash values, which are expressed as: hash ═ h0,h1,h2,...]At this time, the images having the same hash value are classified into one class.
In this embodiment, image features are extracted mainly by a pre-trained ResNet50 model, and ResNet50 is a neural network formed by stacking a series of convolutional layers, and its structure is shown in fig. 2. Each layer other than the layer [3x3maxpool,64] representing the maximum pooling layer represents one convolutional layer, and the convolutional layer represented by [7x7,64/2] represents a convolutional layer having a convolutional kernel size of 7x7, a channel number of 64, and a step size of 2, and the other layers have a step size of 1, and have a total of 50 layers unless otherwise specified.
For each picture, before input, the picture is scaled to 224x224, and after feature extraction, a feature vector of 2048 dimensions is generated.
Then, for each feature vector x, using L2 regularization, after L2 regularization, the feature vector will be scaled to a unit vector in the high-dimensional spherical space.
Then, a projection matrix with a size of 2048xB is randomly generated, where B is the size of the set hash bucket. The projection matrix is globally unique, then matrix multiplication is carried out on the eigenvector and the matrix to obtain a 1xB result vector, and finally the index corresponding to the maximum numerical value in the result vector is taken as the hash value of the eigenvector.
In order to reduce the probability of hash collision, multiple projection matrices may be used for projection to obtain a set of hash values, and the eigenvectors that hold the same hash will be classified into a class, and the classification process is shown in fig. 3. By the image data dividing method, the image characteristics can be effectively extracted, the image data set can be accurately divided, and the probability of Hash collision is reduced.
The foregoing shows and describes the general principles and broad features of the present invention and advantages thereof. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are given by way of illustration of the principles of the present invention, but that various changes and modifications may be made without departing from the spirit and scope of the invention, and such changes and modifications are within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (6)
1. A method for classifying and partitioning an image data set, the method comprising the steps of:
s1, building a pre-training model, randomly generating a projection matrix group R, zooming the image data of the pictures without the classification marks, and importing the image data into the pre-training model;
s2, performing feature extraction on the imported image data without the classified pictures through a pre-training model to generate a feature vector x with a certain dimensionality;
s3, performing L2 regularization on each generated feature vector x by using a regularization formula, and scaling the feature vectors into unit vectors in a high-dimensional spherical space;
s4, for each projection matrix Ri, calculating the matrix by using a Hash function calculation formula to obtain a result vector, and taking an index corresponding to the maximum value in the result vector as a Hash value hi of the feature vector; the hash function calculation formula in step S4 is: hi ═ argmax (x × R)i)；
S5, all the hash values hi calculated by the projection matrix set R in step S4 are combined to obtain a set of hash values and the hash values are used as hashes of the image feature quantities, and at this time, the images with the same hash value are classified into one type.
2. The method for classifying and dividing an image data set according to claim 1, wherein the size of the projection matrix set in step S1 is 2048 × B, where B is the set hash bucket size.
3. The method as claimed in claim 1, wherein the pre-training model in step S1 is ResNet50 model.
4. The method for classifying and dividing an image data set according to claim 1, wherein the size of the scaled image data in step S1 is 224x 224.
5. The method for classifying and dividing an image data set according to claim 1, wherein the regularization formula in the step S3 is:
wherein x isiRepresenting the ith feature in the x vector.
6. The method for classifying and dividing an image data set according to claim 1, wherein the hash of the image feature vector in step S5 is represented by: hash ═ h0,h1,h2,...]。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010722578.3A CN111881312B (en) | 2020-07-24 | 2020-07-24 | Image data set classification and division method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010722578.3A CN111881312B (en) | 2020-07-24 | 2020-07-24 | Image data set classification and division method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111881312A CN111881312A (en) | 2020-11-03 |
| CN111881312B true CN111881312B (en) | 2022-07-05 |
Family
ID=73200251
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010722578.3A Active CN111881312B (en) | 2020-07-24 | 2020-07-24 | Image data set classification and division method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111881312B (en) |
Family Cites Families (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2483851A1 (en) * | 2009-09-29 | 2012-08-08 | NEC Europe Ltd. | Method and system for probabilistic processing of data |
| CN104182538B (en) * | 2014-09-01 | 2017-06-13 | 西安电子科技大学 | Image search method based on semi-supervised Hash |
| US9721190B2 (en) * | 2014-12-19 | 2017-08-01 | Google Inc. | Large-scale classification in neural networks using hashing |
| CN106777038B (en) * | 2016-12-09 | 2019-06-14 | 厦门大学 | A kind of ultralow complexity image search method retaining Hash based on sequence |
| CN106815362B (en) * | 2017-01-22 | 2019-12-31 | 福州大学 | KPCA (Key performance analysis) -based multi-table index image hash retrieval method |
| CN107180079B (en) * | 2017-04-28 | 2020-04-28 | 华南理工大学 | Image retrieval method based on convolutional neural network and tree and hash combined index |
| CN109284411B (en) * | 2017-07-19 | 2022-03-18 | 哈尔滨工业大学深圳研究生院 | Discretization image binary coding method based on supervised hypergraph |
| CN108108657B (en) * | 2017-11-16 | 2020-10-30 | 浙江工业大学 | Method for correcting locality sensitive Hash vehicle retrieval based on multitask deep learning |
| CN108805157B (en) * | 2018-04-11 | 2022-02-18 | 南京理工大学 | Remote sensing image classification method based on partial random supervision discrete hash |
| CN110188227B (en) * | 2019-05-05 | 2022-11-18 | 华南理工大学 | Hash image retrieval method based on deep learning and low-rank matrix optimization |
| CN111177432B (en) * | 2019-12-23 | 2020-11-03 | 北京航空航天大学 | Large-scale image retrieval method based on hierarchical depth hash |
-
2020
- 2020-07-24 CN CN202010722578.3A patent/CN111881312B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN111881312A (en) | 2020-11-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107330127B (en) | Similar text detection method based on text picture retrieval | |
| CN108038122B (en) | Trademark image retrieval method | |
| CN108875076B (en) | Rapid trademark image retrieval method based on Attention mechanism and convolutional neural network | |
| CN110222218B (en) | Image retrieval method based on multi-scale NetVLAD and depth hash | |
| CN109063719B (en) | Image classification method combining structure similarity and class information | |
| CN111340123A (en) | Image score label prediction method based on deep convolutional neural network | |
| CN112100346A (en) | Visual question-answering method based on fusion of fine-grained image features and external knowledge | |
| CN110175221B (en) | Junk short message identification method by combining word vector with machine learning | |
| WO2019137185A1 (en) | Image screening method and apparatus, storage medium and computer device | |
| CN109885796B (en) | Network news matching detection method based on deep learning | |
| CN111898703B (en) | Multi-label video classification method, model training method, device and medium | |
| CN106997379B (en) | Method for merging similar texts based on click volumes of image texts | |
| CN108154156B (en) | Image set classification method and device based on neural topic model | |
| CN114549913B (en) | Semantic segmentation method and device, computer equipment and storage medium | |
| CN110751027B (en) | Pedestrian re-identification method based on deep multi-instance learning | |
| CN110008365B (en) | Image processing method, device and equipment and readable storage medium | |
| CN114064948A (en) | Hash image retrieval method and device based on generalized average pooling strategy | |
| CN110111365B (en) | Training method and device based on deep learning and target tracking method and device | |
| US20230072445A1 (en) | Self-supervised video representation learning by exploring spatiotemporal continuity | |
| CN114882278A (en) | Tire pattern classification method and device based on attention mechanism and transfer learning | |
| CN108304588B (en) | Image retrieval method and system based on k neighbor and fuzzy pattern recognition | |
| CN110210572B (en) | Image classification method, device, storage medium and equipment | |
| CN105844299B (en) | A kind of image classification method based on bag of words | |
| CN111881312B (en) | Image data set classification and division method | |
| CN106529601A (en) | Image classification prediction method based on multi-task learning in sparse subspace |
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 |