CN111639206A - Effective fine image classification method based on optimized feature weight - Google Patents
Effective fine image classification method based on optimized feature weight Download PDFInfo
- Publication number
- CN111639206A CN111639206A CN202010365517.6A CN202010365517A CN111639206A CN 111639206 A CN111639206 A CN 111639206A CN 202010365517 A CN202010365517 A CN 202010365517A CN 111639206 A CN111639206 A CN 111639206A
- Authority
- CN
- China
- Prior art keywords
- image
- data set
- size
- adopted
- network model
- 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 20
- 239000013598 vector Substances 0.000 claims abstract description 39
- 238000012549 training Methods 0.000 claims abstract description 22
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 18
- 238000012360 testing method Methods 0.000 claims abstract description 17
- 238000007781 pre-processing Methods 0.000 claims abstract description 9
- 230000000694 effects Effects 0.000 claims abstract description 6
- 238000011176 pooling Methods 0.000 claims description 20
- 238000004422 calculation algorithm Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 description 5
- 238000000605 extraction Methods 0.000 description 4
- 238000007635 classification algorithm Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012216 screening Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
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
- 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
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an effective fine image classification method based on optimized feature weight, which has high classification accuracy on fine images. The invention discloses an effective fine image classification method based on optimized feature weight, which comprises the following steps: (10) image dataset preprocessing: uniformly cutting and finely cutting the fine images, taking 70% of the preprocessed fine images as a training data set, and taking the rest fine images as a test data set to obtain a final preprocessed image data set; (20) initializing feature weight vectors: initializing the characteristic weight by adopting standard normal distribution; (30) building and training a network model: building a convolutional neural network model, training a characteristic weight vector, optimizing the characteristic weight vector, inputting a test image data set into the trained model and the characteristic weight, testing the model effect of the data set, and outputting a classification result.
Description
Technical Field
The invention belongs to the field of digital image processing and deep learning, and relates to the field of computer vision. In particular to an effective fine image classification method based on optimized feature weight.
Background
The method for classifying the images by utilizing the existing deep convolutional neural network is a method which is used more at present, the calculated amount of the network is determined by the size of input image data of the deep convolutional neural network, and the characteristic information needs to be continuously discarded by the traditional network, so that the information is insufficient during later analysis. There is a research to propose a feature extraction model that can sufficiently retain image features, and is used for performing lossless feature extraction operation on an input image with an arbitrary resolution. For example, document 1(CN 1106596553A, "a feature extraction model and a feature extraction method capable of sufficiently retaining image features")
For a fine image, the image is similar on the whole, and the image can be generally classified through some partial features, but for a deep convolutional neural network, the learning mode is autonomous learning, which features influence the classification result cannot be determined, and the features influencing the classification can be found out through a continuous learning process. There is a research to propose an improved fine image classification method for center loss, which adds a constraint term for calculating the nearest center distance of other classes in the loss function based on center loss, and the constraint term can increase the distance between classes. See document 2 (Row. Fine image Classification Algorithm based on deep learning study [ D ])
Current methods for improving the accuracy of image classification algorithms are generally considered from the following aspects: data set preprocessing, model adjustment, loss function optimization, and the like. In most image classification tasks, the difference of the internal features of the images is obvious, and the classification can achieve higher accuracy by using common convolutional neural networks such as AlexNet, VGG16 and ResNet, but the accuracy which can be obtained by using a general convolutional neural network model to perform a fine image classification task is lower.
Disclosure of Invention
The invention aims to provide an effective fine image classification method based on optimized feature weight, which has high classification accuracy on fine images.
The technical solution for realizing the invention is as follows:
an effective fine image classification method based on optimized feature weight comprises the following steps:
(10) image dataset preprocessing: uniformly cutting and finely cutting the fine images, taking 70% of the preprocessed fine images as a training data set, and taking the rest fine images as a test data set to obtain a final preprocessed image data set;
(20) initializing feature weight vectors: initializing the characteristic weight by adopting standard normal distribution;
(30) building and training a network model: building a convolutional neural network model, training a characteristic weight vector, optimizing the characteristic weight vector, inputting a test image data set into the trained model and the characteristic weight, testing the model effect of the data set, and outputting a classification result.
Compared with the prior art, the invention has the following remarkable advantages:
the accuracy of fine image classification can be improved, mainly because:
(1) the difference between the fine image and the common image is fully considered, and the fine image is similar on the whole and different in partial characteristics. Segmenting the fine image, and classifying the image more accurately by researching partial area of the image
(2) A simple layer of artificial neural network is added behind the convolutional neural network model, the increased characteristic weight indicates that the decisive action of each divided image area on the whole image classification task is different, and the characteristic decisive action is obvious and is large, otherwise, the characteristic decisive action is small.
Drawings
Fig. 1 is a main flow chart of an effective fine image classification method based on optimized feature weights according to the present invention.
FIG. 2 is a flow chart of the network model building and training steps in FIG. 1.
Detailed Description
The invention is described in further detail below with reference to the figures and the embodiments.
As shown in FIG. 1, the effective fine image classification method based on optimized feature weight of the present invention comprises the following steps:
(10) image dataset preprocessing: uniformly cutting and finely cutting the fine images, taking 70% of the preprocessed fine images as a training data set, and taking the rest fine images as a test data set to obtain a final preprocessed image data set;
preferably, the (10) image dataset preprocessing step comprises:
(11) image uniform cutting: uniformly cutting the image data set into images m × m with consistent length and width, wherein m is the side length of the images;
(12) fine image cutting: finely cutting the images after the uniform cutting, cutting each image by adopting a step length of stride and a cutting frame size of N X N from left to right and from top to bottom to obtain N X N image data, wherein each image data set is XijWhere N is X after preprocessing of the image datasetijNumber, XijThe image data with coordinates (i, j) at the upper left corner and size n x n of each category image is intercepted. Wherein N is the side length of the cutting frame, and N is (m-N)/stride +1, 0<i<m,0<j<m,0<n<m。
(20) Initializing feature weight vectors: initializing the characteristic weight by adopting standard normal distribution;
the step (20) of initializing the feature weight vector specifically comprises the following steps:
initializing a characteristic weight vector W, wherein the W is a two-dimensional vector, the initialization mode is standard normal distribution, and the value of the characteristic weight vector satisfiesx is a parameter of each feature weight vector.
(30) Building and training a network model: building a convolutional neural network model, training a characteristic weight vector, optimizing the characteristic weight vector, inputting a test image data set into the trained model and the characteristic weight, testing the model effect of the data set, and outputting a classification result.
In the step of (30) network model building and training, the convolutional neural network model built in the network model step is VGG16, and for the input image data n × n, the built VGG16 model is:
block1, 2 convolution layers, each convolution kernel is 3 x 3 in size, the number of channels is 64, then a pooling layer is adopted, maxpool is adopted, and the size of the pooling layer is 2 x 2;
block 2: 2 convolution layers, wherein the size of each convolution kernel is 3 x 3, the number of channels is 128, then a pooling layer is adopted, maxpool is adopted, and the size of the pooling layer is 2 x 2;
block 3: 3 convolution layers, wherein the size of each convolution kernel is 3 x 3, the number of channels is 256, then a pooling layer is adopted, maxpool is adopted, and the size of the pooling layer is 2 x 2;
block 4: 3 convolution layers, wherein the size of each convolution kernel is 3 x 3, the number of channels is 512, then a pooling layer is adopted, maxpool is adopted, and the size of the pooling layer is 2 x 2;
block 5: 3 convolution layers, wherein the size of each convolution kernel is 3 x 3, the number of channels is 512, then a pooling layer is adopted, maxpool is adopted, and the size of the pooling layer is 2 x 2;
block 6: 3 full-connected layers, wherein the weight number of the first layer is 4096, the weight number of the second layer is 4096, the weight number of the 3 rd layer is 1000, and finally, a classification result s x 1 is output through softmax;
where s is the number of categories of the image classification task.
Preferably, as shown in fig. 2, the (30) network model building and training step includes:
(31) building a network model: building a VGG16 convolutional neural network model;
(32) training the feature weight vector: training a feature weight vector by taking an output result of the convolutional neural network as an input;
respectively inputting the preprocessed N × N training data sets into a VGG16 network model, setting iteration times t and a screening rate theta, outputting a classification result s × 1 of the convolutional neural network after the network is trained for t times, sorting the classification results, and screening out the theta × N2And splicing R classification results into a two-dimensional vector A, wherein the structure of the A is s R, and R is (1-theta) N2。
(33) Optimizing the feature weight vector: reducing the loss function value according to a back propagation algorithm to optimize the feature weight vector;
multiplying the initialized feature vector W by the vector A and adding the multiplied feature vector W to an offset B, namely Y-A-W + B, wherein the feature weight W is of the structure R1 and the offset B-mu1…μk…μs]Randomly initialized to a number between (0,0.5), where 1<k<s, the structure of the Y vector is s 1;
passing the output result Y and the actual tag value C through a softmax cross entropy loss functionOptimizing a Loss function value by adopting an Adam optimizer, continuously changing the characteristic weight value W and the bias value B through a back propagation algorithm until the Loss function value is smaller than a threshold value Q, and stopping updating operation;
(34) testing data set to check model effect: and inputting the test data set into the trained VGG16 network model and the feature weight vector to obtain a model classification result.
And inputting the test data set into a trained VGG16 convolutional neural network model, inputting the output result into the updated feature weight and bias again for operation, and finally obtaining a classification result.
The fine image data processed by the method is image data which is difficult to classify, the image data in the data set are integrally similar, and the classified features are difficult to find and identify. Therefore, by cutting the image data and trying to classify the local information of each image, the difference of the fine image in the partial image area is large, and the difference of the fine image in a certain area is small.
Claims (5)
1. An effective fine image classification method based on optimized feature weight is characterized by comprising the following steps:
(10) image dataset preprocessing: uniformly cutting and finely cutting the fine images, taking 70% of the preprocessed fine images as a training data set, and taking the rest fine images as a test data set to obtain a final preprocessed image data set;
(20) initializing feature weight vectors: initializing the characteristic weight by adopting standard normal distribution;
(30) building and training a network model: building a convolutional neural network model, training a characteristic weight vector, optimizing the characteristic weight vector, inputting a test image data set into the trained model and the characteristic weight, testing the model effect of the data set, and outputting a classification result.
2. The efficient fine image classification method according to claim 1, characterized in that the (10) image dataset preprocessing step comprises:
(11) image uniform cutting: uniformly cutting the image data set into images m × m with consistent length and width, wherein m is the side length of the images;
(12) fine image cutting: finely cutting the images after the uniform cutting, cutting each image by adopting a step length of stride and a cutting frame size of N X N from left to right and from top to bottom to obtain N X N image data, wherein each image data set is XijWhere N is X after preprocessing of the image datasetijNumber, XijThe image data with coordinates (i, j) at the upper left corner and size n x n of each category image is intercepted. Wherein N is the side length of the cutting frame, and N is (m-N)/stride +1, 0<i<m,0<j<m,0<n<m。
3. The efficient fine image classification method according to claim 1, characterized in that the (20) initializing feature weight vectors step is specifically:
4. The efficient fine image classification method according to claim 1, wherein in the (30) network model building and training step, the convolutional neural network model built in the network model step is VGG16, and for the input image data n x n, the built VGG16 model is as follows:
block1, 2 convolution layers, each convolution kernel is 3 x 3 in size, the number of channels is 64, then a pooling layer is adopted, maxpool is adopted, and the size of the pooling layer is 2 x 2;
block 2: 2 convolution layers, wherein the size of each convolution kernel is 3 x 3, the number of channels is 128, then a pooling layer is adopted, maxpool is adopted, and the size of the pooling layer is 2 x 2;
block 3: 3 convolution layers, wherein the size of each convolution kernel is 3 x 3, the number of channels is 256, then a pooling layer is adopted, maxpool is adopted, and the size of the pooling layer is 2 x 2;
block 4: 3 convolution layers, wherein the size of each convolution kernel is 3 x 3, the number of channels is 512, then a pooling layer is adopted, maxpool is adopted, and the size of the pooling layer is 2 x 2;
block 5: 3 convolution layers, wherein the size of each convolution kernel is 3 x 3, the number of channels is 512, then a pooling layer is adopted, maxpool is adopted, and the size of the pooling layer is 2 x 2;
block 6: 3 full-connected layers, wherein the weight number of the first layer is 4096, the weight number of the second layer is 4096, the weight number of the 3 rd layer is 1000, and finally, a classification result s x 1 is output through softmax;
where s is the number of categories of the image classification task.
5. The efficient fine image classification method according to claim 1, characterized in that the (30) network model building and training step comprises:
(31) building a network model: building a VGG16 convolutional neural network model;
(32) training the feature weight vector: training a feature weight vector by taking an output result of the convolutional neural network as an input;
(33) optimizing the feature weight vector: reducing the loss function value according to a back propagation algorithm to optimize the feature weight vector;
(34) testing data set to check model effect: and inputting the test data set into the trained VGG16 network model and the feature weight vector to obtain a model classification result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010365517.6A CN111639206A (en) | 2020-04-30 | 2020-04-30 | Effective fine image classification method based on optimized feature weight |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010365517.6A CN111639206A (en) | 2020-04-30 | 2020-04-30 | Effective fine image classification method based on optimized feature weight |
Publications (1)
Publication Number | Publication Date |
---|---|
CN111639206A true CN111639206A (en) | 2020-09-08 |
Family
ID=72329005
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010365517.6A Pending CN111639206A (en) | 2020-04-30 | 2020-04-30 | Effective fine image classification method based on optimized feature weight |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111639206A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2022141511A1 (en) * | 2020-12-31 | 2022-07-07 | 深圳市优必选科技股份有限公司 | Image classification method, computer device, and storage medium |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109285112A (en) * | 2018-09-25 | 2019-01-29 | 京东方科技集团股份有限公司 | Image processing method neural network based, image processing apparatus |
-
2020
- 2020-04-30 CN CN202010365517.6A patent/CN111639206A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109285112A (en) * | 2018-09-25 | 2019-01-29 | 京东方科技集团股份有限公司 | Image processing method neural network based, image processing apparatus |
Non-Patent Citations (1)
Title |
---|
吕恩辉: "基于卷积神经网络的图像分类研究", 《中国博士学位论文全文数据库 (信息科技辑)》 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2022141511A1 (en) * | 2020-12-31 | 2022-07-07 | 深圳市优必选科技股份有限公司 | Image classification method, computer device, and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US20190228268A1 (en) | Method and system for cell image segmentation using multi-stage convolutional neural networks | |
US20210049423A1 (en) | Efficient image classification method based on structured pruning | |
CN108985317B (en) | Image classification method based on separable convolution and attention mechanism | |
CN110334765B (en) | Remote sensing image classification method based on attention mechanism multi-scale deep learning | |
CN110197205B (en) | Image identification method of multi-feature-source residual error network | |
CN111079795B (en) | Image classification method based on CNN (content-centric networking) fragment multi-scale feature fusion | |
CN110826684B (en) | Convolutional neural network compression method, convolutional neural network compression device, electronic device, and medium | |
CN110322445B (en) | Semantic segmentation method based on maximum prediction and inter-label correlation loss function | |
CN108614997B (en) | Remote sensing image identification method based on improved AlexNet | |
CN111414942A (en) | Remote sensing image classification method based on active learning and convolutional neural network | |
CN112163628A (en) | Method for improving target real-time identification network structure suitable for embedded equipment | |
CN107392919B (en) | Adaptive genetic algorithm-based gray threshold acquisition method and image segmentation method | |
CN111696101A (en) | Light-weight solanaceae disease identification method based on SE-Inception | |
CN109033978B (en) | Error correction strategy-based CNN-SVM hybrid model gesture recognition method | |
CN112101364B (en) | Semantic segmentation method based on parameter importance increment learning | |
CN111145145B (en) | Image surface defect detection method based on MobileNet | |
CN109949200B (en) | Filter subset selection and CNN-based steganalysis framework construction method | |
CN112861718A (en) | Lightweight feature fusion crowd counting method and system | |
CN110991554B (en) | Improved PCA (principal component analysis) -based deep network image classification method | |
CN106503047A (en) | A kind of image reptile optimization method based on convolutional neural networks | |
CN113297936A (en) | Volleyball group behavior identification method based on local graph convolution network | |
CN110956201A (en) | Image distortion type classification method based on convolutional neural network | |
CN114882278A (en) | Tire pattern classification method and device based on attention mechanism and transfer learning | |
CN114429298A (en) | PCB order batch optimization method based on BP neural network | |
CN111639206A (en) | Effective fine image classification method based on optimized feature weight |
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: 20200908 |