CN104281843A - Image recognition method and system based on self-adaptive features and classification model selection - Google Patents

Abstract

The present invention provides an image recognition method and system based on adaptive features and classification model selection, which fully considers the recognition of different types of images, is not limited to the matching of a single image feature, and is also different from the combined feature matching method, and has a relatively High flexibility. The present invention pre-sets a relatively comprehensive feature extraction method for the image recognition retrieval model, and sets a corresponding classifier for each feature, and at the same time determines the method for calculating the distance between image features, and determines the distance between the features in the feature library according to the result of the distance calculation. Identify the feature with the closest feature distance of the image, and adaptively designate a suitable classifier for the image to be classified and recognized to perform feature matching and class discrimination, so as to improve the image recognition rate. When the image to be recognized is affected by various adverse factors, use The same recognition model has a good recognition effect, which can solve the problem of reduced matching performance of the recognition model when the type of the target image changes.

Description

Image recognition method and system based on adaptive feature and classification model selection

technical field

The invention relates to the technical fields of image processing, pattern recognition and machine learning, in particular to an image recognition method and system based on adaptive features and classification model selection.

Background technique

At present, image classification and recognition methods are mostly based on image features for matching, recognition, and retrieval, that is, to design an algorithm model with machine learning methods, and use a certain number of samples of image databases to extract features from a certain feature extraction method. Training, so that it has a certain cognitive ability; after that, the model inputs a sample image of an unknown category, that is, the test image, and matches the unknown image with the known image database according to similarity and other criteria, so as to determine the quality of the input test image. identity.

The existing image classification and recognition methods mainly have the following shortcomings:

(1) It is necessary to determine the appropriate image feature extraction method in advance. Since the image of the category to be determined has many unknown factors, such as missing parts of the image, camera angle deflection, etc., the preset feature extraction method is not suitable for all samples. .

(2) Different feature extraction methods must be combined with corresponding classification recognition algorithms to reflect better recognition results. To obtain accurate recognition rates, the algorithm model is more complicated.

Contents of the invention

The purpose of the present invention is to provide an image recognition method and system based on adaptive features and classification model selection, which can solve the problem of reduced matching performance of the recognition model when the type of the target image changes.

In order to solve the above problems, the present invention provides an image recognition method based on adaptive features and classification model selection, including:

An image sample database for training and testing is set, and the image sample database is divided into several subcategories, and the image sample data in each subcategory has one or more common features;

Determine the extraction method of each feature, and determine the corresponding classifier for each feature extraction method;

For the image sample data in all subcategories, extract features according to the extraction methods of all features to form a feature library, and perform consistent processing on all the features in the feature library;

Each feature in the feature library is trained with all the classifiers in turn, and some images are randomly selected from the feature library to verify the recognition rate of the classifier, and the verification results for each subclass according to the recognition rate Determine a feature with the highest recognition rate and the corresponding classifier;

Use all feature extraction methods to extract the corresponding features of the image to be recognized, and perform consistent processing on the extracted features;

Determine the distance between the features of the image to be recognized and each feature in the feature library according to a preset calculation distance method, and determine the feature in the feature library that is closest to the feature of the image to be recognized according to the result of the distance calculation;

Determine the distance between the feature of the image to be recognized and each feature in the feature library according to a preset calculation distance method, and determine the feature in the feature library that is closest to the feature of the image to be recognized according to the result of the calculation distance;

Determine the classifier corresponding to the feature according to the feature closest to the feature of the image to be recognized, determine the corresponding subclass according to the determined corresponding classifier, use the determined corresponding classifier to retrieve and identify in the corresponding subclass Image sample data matched with the image to be recognized.

Further, in the above method, the extraction method is the LBP method.

Further, in the above method, the extraction method is the SIFT method.

According to another aspect of the present invention, an image recognition system based on adaptive features and classification model selection is provided, including:

The division module is used to set the image sample database for training and testing, and divides the image sample database into several subcategories, and the image sample data in each subcategory has one or more common characteristics;

A determination module is used to determine the extraction method of each feature, and determine a corresponding classifier for each feature extraction method;

The first extraction module is used to extract features from image sample data in all subcategories according to all feature extraction methods to form a feature library, and perform consistent processing on all features in the feature library, so that different features can be used in the same standards for comparison;

The training and selection module is used to train each feature in the feature library with all classifiers in turn, and randomly select some images from the feature library to verify the recognition rate of the classifier, according to the recognition rate According to the verification results, select a feature with the highest recognition rate and a corresponding classifier for each subclass;

The second extraction module is used to extract corresponding features using all feature extraction methods of the image to be identified, and perform consistent processing on the extracted features, so that different features can be compared with a unified criterion;

A distance module, configured to determine the distance between the feature of the image to be recognized and each feature in the feature library according to a preset calculation distance method, and determine the feature distance between the feature library and the image to be recognized according to the result of the calculation distance most recent features;

The matching module is used to determine the corresponding classifier according to the feature closest to the feature of the image to be recognized, determine the corresponding subclass according to the determined corresponding classifier, and use the determined corresponding classifier to retrieve and identify in the corresponding subclass Output image sample data that matches the image to be recognized.

Further, in the above system, the extraction method is the LBP method.

Further, in the above system, the extraction method is a SIFT method.

Compared with the prior art, the model of the present invention fully considers the recognition of different types of images, is not limited to the matching of a single image feature, and is also different from the combined feature matching method. It has high flexibility and is an image recognition retrieval model. Set a more comprehensive feature extraction method in advance, and set a corresponding classifier for each feature; at the same time, calculate the distance method and determine the feature with the closest distance to the feature of the image to be recognized in the feature library according to the result of the calculation distance, for the classifier to be classified The recognized image adaptively specifies the appropriate classifier for feature matching and class discrimination to improve the image recognition rate, and can adaptively select the correct feature extraction method and corresponding recognition retrieval for various images to be determined method, when the image to be recognized is affected by various adverse factors, using the same recognition model has a better recognition effect, has high practicability, and can solve the matching performance of the recognition model when the type of the target image changes. The reduced problem has the advantages of reliability, high real-time performance, and strong robustness, and can provide reliable support for intelligent monitoring and other fields.

Description of drawings

Fig. 1 is a flow chart of an image recognition method based on adaptive features and classification model selection according to an embodiment of the present invention;

Fig. 2 is a block diagram of an image recognition system based on adaptive feature and classification model selection according to an embodiment of the present invention.

Detailed ways

In order to make the above objects, features and advantages of the present invention more comprehensible, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

Embodiment one

As shown in Figure 1, the present invention provides a kind of image recognition method based on adaptive feature and classification model selection, comprising:

Step S1, setting an image sample database for training and testing, and dividing the image sample database into several subcategories, and the image sample data in each subcategory has one or more common features;

Step S2, determine the extraction method of each feature, and determine the corresponding classifier for each feature extraction method;

Step S3, extracting features from the image sample data in all subcategories according to all feature extraction methods to form a feature library, and performing consistent processing on all the features in the feature library, so that different features can be compared with a unified criterion;

Step S4, use all classifiers to train each feature in the feature library in turn, and randomly select some images from the feature library to verify the recognition rate of the classifier. The verification result according to the recognition rate is Each subclass determines a feature with the highest recognition rate and a corresponding classifier;

Step S5, using all feature extraction methods to extract corresponding features from the image to be recognized, and performing consistent processing on the extracted features, so that different features can be compared with a unified criterion;

Step S6: Determine the distance between the feature of the image to be recognized and each feature in the feature library according to the preset calculation distance method, and determine the feature library with the closest distance to the feature of the image to be recognized according to the distance calculation result feature;

Step S7, determine the classifier corresponding to the feature according to the feature closest to the feature of the image to be recognized, determine the corresponding subclass according to the determined corresponding classifier, and use the determined corresponding classifier in the corresponding subclass Retrieve and identify image sample data matching the image to be recognized.

Preferably, the extraction method is an LBP method (Local Binary Patterns).

Preferably, the extraction method is a SIFT method (scale-invariant feature transform, Scale-invariantfeature transform).

The model of this embodiment fully considers the recognition of different types of images, and is not limited to the matching of a single image feature, and is also different from the combined feature matching method, and has higher flexibility. Pre-set a relatively comprehensive feature extraction method for the image recognition retrieval model, and set a corresponding classifier for each feature; at the same time determine the method of calculating the distance between image features and determine the distance between the feature library and the image to be recognized according to the result of the calculation distance. The feature with the closest feature distance is used to adaptively designate a suitable classifier for feature matching and class discrimination for the image to be classified and recognized, so as to improve the image recognition rate, and can adaptively provide various categories to be determined within the test range Select the correct feature extraction method and the corresponding recognition retrieval method for the image. When the image to be recognized is affected by various adverse factors, using the same recognition model will have a better recognition effect and high practicability, which can solve the problem of The matching performance of the recognition model decreases when the target image type changes, etc. It has the advantages of reliability, high real-time performance, and strong robustness, and can provide reliable support for intelligent monitoring and other fields.

Embodiment two

As shown in Figure 2, the present invention also provides another image recognition system based on adaptive features and classification model selection, including:

Divide module 1, be used to set the image sample database for training and testing, and divide described image sample database into several subcategories, image sample data has one or more common features in each subcategory;

Determining module 2, used to determine the extraction method of each feature, and determine the corresponding classifier for the extraction method of each feature;

The first extraction module 3 is used to extract features from the image sample data in all subcategories according to the extraction methods of all features to form a feature library, and perform consistent processing on all the features in the feature library to make different features available. compared with the same criteria;

Training and selection module 4 is used to train each feature in the feature library with all classifiers in turn, and randomly select part of the images from the feature library to verify the recognition rate of the classifier, according to the identification The verification result of the rate selects a feature with the highest recognition rate and a corresponding classifier for each subclass;

The second extraction module 5 is used to extract corresponding features using all feature extraction methods of the image to be identified, and perform consistency processing on the extracted features, so that different features can be compared with a unified criterion;

The distance module 6 is used to determine the distance between all the features of the image to be recognized and each feature in the feature library according to the preset calculation distance method, and determine the distance between the image to be recognized in the feature library and the image to be recognized according to the result of the distance calculation. The feature with the closest feature distance;

Matching module 7, for determining the classifier corresponding to the feature according to the feature closest to the feature distance of the image to be recognized, determining the corresponding subclass according to the determined corresponding classifier, using the determined corresponding classifier in the corresponding In the subclass, the image sample data that matches the image to be recognized is identified.

Preferably, the extraction method is an LBP method (Local Binary Patterns).

Preferably, the extraction method is a SIFT method (scale-invariant feature transform, Scale-invariantfeature transform).

In summary, the model of the present invention fully considers the recognition of different types of images, is not limited to the matching of single image features, and is also different from the combination of feature matching methods. It has high flexibility and is preset for the image recognition retrieval model. A relatively comprehensive feature extraction method, and set a corresponding classifier for each feature, and at the same time determine the method of calculating the distance between image features and determine the feature in the feature library with the closest feature distance to the feature of the image to be recognized according to the result of the calculation distance. In order to adaptively designate a suitable classifier for feature matching and class discrimination for the image to be classified and recognized, so as to improve the image recognition rate, it can adaptively select the correct feature extraction for a variety of images to be determined within the test range method and the corresponding recognition and retrieval method, when the image to be recognized is affected by various adverse factors, using the same recognition model has a better recognition effect, has high practicability, and can solve the situation that the target image type changes, etc. It has the advantages of reliability, high real-time performance, and strong robustness, and can provide reliable support for intelligent monitoring and other fields.

Each embodiment in this specification is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same and similar parts of each embodiment can be referred to each other. As for the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and for related parts, please refer to the description of the method part.

Professionals can further realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software or a combination of the two. In order to clearly illustrate the possible For interchangeability, in the above description, the composition and steps of each example have been generally described according to their functions. Whether these functions are executed by hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the present invention.

Obviously, those skilled in the art can make various changes and modifications to the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and equivalent technologies thereof, the present invention also intends to include these modifications and variations.

Claims (6)

1., based on the image-recognizing method that self-adaptive features and disaggregated model are selected, it is characterized in that, comprising:

Arrange the image sample data storehouse for training and testing, and described image sample data storehouse is divided into several subclasses, in each subclass, image sample data has one or more common features;

Determine the extracting method of each feature, and determine corresponding sorter for the extracting method of each feature;

According to institute's characteristic extracting method, feature composition characteristic storehouse is extracted respectively to the image sample data in all subclasses, and consistency treatment is carried out to features all in described feature database;

Each feature in described feature database is trained in turn with all sorters, and Stochastic choice parts of images, in order to verify the discrimination of sorter, is the sorter of a kind of feature that each subclass determination discrimination is the highest and correspondence according to the result of discrimination from described feature database;

Treat the characteristic extracting method of recognition image institute and extract characteristic of correspondence, and consistency treatment is carried out to the feature extracted;

According to the distance of each feature in feature database described in all characteristic sum that the calculating distance method preset determines described image to be identified, according to feature nearest with the characteristic distance of image to be identified in the result determination feature database that distance calculates;

According to the distance of each feature in feature database described in the characteristic sum that the calculating distance method preset determines described image to be identified, according to calculating feature nearest with the characteristic distance of image to be identified in the result determination feature database of distance;

The sorter corresponding with this feature is determined according to the feature nearest with the characteristic distance of image to be identified, determine corresponding subclass according to the sorter of the correspondence determined, use sorter retrieval in the subclass of correspondence of the correspondence determined to go out the image sample data with images match to be identified.

2., as claimed in claim 1 based on the image-recognizing method that self-adaptive features and disaggregated model are selected, it is characterized in that, described extracting method is LBP method.

3., as claimed in claim 1 based on the image-recognizing method that self-adaptive features and disaggregated model are selected, it is characterized in that, described extracting method is SIFT method.

4., based on the image identification system that self-adaptive features and disaggregated model are selected, it is characterized in that, comprising:

Divide module, for arranging the image sample data storehouse for training and testing, and described image sample data storehouse is divided into several subclasses, in each subclass, image sample data has one or more common features;

Determination module, for determining the extracting method of each feature, and determines corresponding sorter for the extracting method of each feature;

First extraction module, for extracting feature composition characteristic storehouse according to the characteristic extracting method of institute respectively to the image sample data in all subclasses, and carries out consistency treatment to features all in described feature database;

Training and selection module, for training in turn with all sorters each feature in described feature database, and Stochastic choice parts of images, in order to verify the discrimination of sorter, is the sorter of a kind of feature that each subclass selective recognition rate is the highest and correspondence according to the result of discrimination from described feature database;

Second extraction module, extracts characteristic of correspondence for treating all feature extracting methods of recognition image, and carries out consistency treatment to the feature extracted;

Spacing module, for determine described image to be identified according to the calculating distance method preset all characteristic sum described in the distance of each feature in feature database, according to feature nearest with the characteristic distance of image to be identified in the result determination feature database that distance calculates;

Matching module, for determining the sorter corresponding with this feature according to the feature nearest with the characteristic distance of image to be identified, determine corresponding subclass according to the sorter of the correspondence determined, use sorter retrieval in the subclass of correspondence of the correspondence determined to go out the image sample data with images match to be identified.

5., based on the image identification system that self-adaptive features and disaggregated model are selected, it is characterized in that, described extracting method is LBP method.

6., based on the image identification system that self-adaptive features and disaggregated model are selected, it is characterized in that, described extracting method is SIFT method.