KR20170109304A - Method for parallel learning of cascade classifier by object recognition - Google Patents

Abstract

The present invention relates to a parallel learning method of a cascade classifier for recognizing an object. According to an embodiment of the present invention, the parallel learning method for recognizing an object which is used to perform object recognition on an image, comprises: a learning preparing step for extracting feature information on a learning image to be input as learning data to a plurality of unit classifiers; a positive learning step of inputting a negative sample to be used with at least one positive sample among the extracted feature information, into the unit classifiers to perform a learning process, and to perform an evaluation so as to determine a result thereof, and then determining priority for the unit classifiers on the basis of information on the evaluation result; and a negative learning step of sequentially inputting at least one negative sample among the extracted feature information into the unit classifiers in accordance with the determined priority to perform a learning process, and to perform an evaluation so as to determine the same as a final unit classifier when a desired result is derived, and when the desired result is not derived, reconstructing the negative sample used in the current turn into a negative sample to be input in the next turn, to input the sample to the next unit classifiers so as to repeat the learning process.

Description

{METHOD FOR PARALLEL LEARNING OF CASCADE CLASSIFIER BY OBJECT RECOGNITION FOR Cascade Classifier for Object Recognition}

The present invention relates to a parallel learning method of a cascade classifier for object recognition, and more particularly, to a learning method in a cascade classifier used for classifying an image or an image in object recognition processing.

Techniques for recognizing objects from images have been actively studied in various fields such as image processing, pattern recognition, computer vision and neural networks, and have many commercial and legal applications.

Recently, a lot of digital images have been circulated on the Internet, since photos and images taken by digital cameras and smart phones have been shared with social network services (SNS), blogs, and community sites. As a result, analytical techniques for images or images have attracted a great deal of attention, and recent studies have been actively conducted

This image analysis technique analyzes one or more objects included in an image and recognizes, classifies and judges the type, personality, and characteristics of each object. Such image analysis techniques may vary depending on the method used, but are usually performed through pre-processing, feature extraction, object learning, and object classification.

Among the image analysis techniques described above, the cascade classifier shows a high classification performance in detecting human faces or detecting pedestrians. The configuration of such a cascade classifier consists of sequentially connecting a plurality of unit classifiers and judging them as objects to be classified when the result values derived from the classifiers are all identical. That is, when the resultant values in any one of the sequentially connected plurality of unit classifiers do not match, it is determined that the object is not the object to be classified.

The cascaded classifier with this feature is designed to reduce the cases of false positive (actually false, but true judgment) in the learning process to include cases judged as false positives in the previous classifier as inputs of the next classifier . Therefore, there is a dependency on the learning data between each unit classifier. There is a problem that it is impossible to simultaneously learn a plurality of unit classifiers sequentially connected due to a structure having a dependency on learning data between unit classifiers. For this reason, as the number of unit classifiers increases, the learning time also increases, which leads to a problem that the time required for final image analysis increases.

In addition, when parallel processing using a GPU commonly used in recent years or cluster processing requires a large number of computers to perform image analysis processing rapidly, parallel processing can not be performed in the learning procedure as described above, The shortening effect can not be realized.

The object of the present invention is to provide an object recognition method capable of improving the sequential learning and processing procedure between unit classifiers in a learning process of a cascade classifier used in an image or image analysis method, The purpose of this paper is to provide a parallel learning method of a cascade classifier.

Other objects and advantages of the present invention will be described hereinafter and will be understood by the embodiments of the present invention. Further, objects and advantages of the present invention can be realized by the means and the combination shown in the appended claims.

In order to achieve the above object, a parallel learning method of a cascade classifier for object recognition of the present invention is a learning processing method in a cascade classifier used for object recognition processing on an image or an image, A learning preparation step of extracting feature information on a learning image to be input as data; A negative sample used in common with at least one positive sample of the extracted feature information is input to a plurality of unit classifiers, learning is performed, the evaluation is performed to determine the result, and then the plurality A positive learning step of determining the priority of the unit classifier of And at least one negative sample of the extracted feature information is sequentially input to the unit classifier according to the determined priority, and the learning is performed. When the desired result is obtained, the final unit classifier is determined. A negative learning step of reconstructing a negative sample to be input next to a negative sample to be input next and a negative learning step of repeating the learning in the next step of the unit classifier, and a method of parallel learning of a cascade classifier for object recognition .

According to the present invention, in the parallel learning method of a cascade classifier for object recognition of the present invention, a case where there is no data dependency between unit classifiers in a learning process of a cascade classifier is first learned, and a dependent part is learned sequentially, It is possible to reduce the time required for the learning process because the parallel processing of the learning can be performed for a certain portion between the unit classifiers.

Furthermore, it is possible to perform a partial partial parallel processing between multiple unit classifiers of a cascade classifier, and it is possible to utilize a parallel processing system using a cluster system or a GPU, thereby reducing not only the processing time of the learning process but also the overall image recognition and analysis processing time It is effective.

In addition, in constructing the classifier, it may happen that the intended performance does not come out due to an over-sum or the like. Therefore, it is possible to randomly configure a plurality of classifiers in addition to one classifier which is determined to be optimum, thereby improving the accuracy of the cascade classifier as a whole Provides a possible effect.

1 is a block diagram of a cascade classifier in which a parallel learning method of a cascade classifier for object recognition is performed according to an embodiment of the present invention.
2 is a diagram illustrating a learning process of a cascade classifier according to another embodiment of the present invention.
3 is a diagram illustrating a procedure of a parallel learning method of a cascade classifier for object recognition according to an embodiment of the present invention.
4 is a diagram illustrating a process of performing a parallel learning method of a cascade classifier for object recognition according to an embodiment of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention and the manner of achieving them will become apparent with reference to the embodiments described in detail below with reference to the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. And is provided to fully convey the scope of the invention to those skilled in the art, and the present invention is defined by the claims. It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. In the present specification, the singular form includes plural forms unless otherwise specified in the specification. It is noted that " comprises, " or "comprising," as used herein, means the presence or absence of one or more other components, steps, operations, and / Do not exclude the addition.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the drawings, the same reference numerals are given to the same or similar components, and in the following description of the present invention, Detailed explanations of the detailed description will be omitted when the gist of the present invention can be obscured.

1 is a block diagram of a cascade classifier in which a parallel learning method of a cascade classifier for object recognition is performed according to an embodiment of the present invention.

Referring to FIG. 1, the construction and processing of a cascade classifier in which a parallel learning method of a cascade classifier for object recognition according to an embodiment of the present invention is performed will be described.

As shown in the drawing, the cascade classifier is configured such that a plurality of unit classifiers 101 are sequentially connected, and classification and determination processing is sequentially performed in each unit classifier 101. In each of the unit classifiers 101 (103, Object) only when it is determined that the target object is the target object to be detected.

At this time, when any one of the unit sorters 101 determines that the object 102 is a non-object rather than a target object to be detected, it is determined that the object is not a corresponding object in the intermediate process up to that point. It is not necessary to perform the classification and judgment processing in the unit classifiers of FIG. Therefore, there is an advantage that the classification result of the object can be derived quickly as a whole.

2 is a diagram illustrating a learning process of a cascade classifier according to another embodiment of the present invention.

Here, in order to learn the cascade classifier, a process of learning a plurality of unit classifiers included in the cascade classifier is sequentially performed. A learning process for each unit classifier will be described.

Referring to FIG. 2, the learning process of the cascade classifier according to another embodiment of the present invention includes a procedure of preparing and receiving a learning image or learning image data to be used for learning, And a preparation process 201 including a process of extracting feature or descriptor information from the loaded image or image. At this time, the feature information to be extracted includes various characteristics used in the analysis before the image recognition such as Scale Invariant Feature Transform (SIFT), Histogram of Oriented Gradient (HOG), Local Binary Pattern (LBP), Edge histogram, Information may be included.

When the learning data is prepared in this manner, a procedure for performing the unit learning 202 is performed using the learning data. In the unit learning step 202, learning is performed using one or a plurality of pieces of extracted feature information. Learning performed in the unit learning 202 is a process of adjusting a parameter so that a predetermined output is output to a predetermined input. The classifier includes a decision tree, a support vector machine (SVM), a neural network Neural Network) can be used. After performing the learning process, the classifier determines an input value that is determined to be similar according to the adjusted parameter as an object. In the unit learning process 202, a positive sample and a negative sample are input together.

When the unit learning 202 procedure is completed, an evaluation 203 procedure is performed. At this time, evaluation is performed using a classifier parameter that is a result of learning in the unit learning 202 procedure. Preferably, the input image used in the evaluation step (203) is not used for learning.

After the evaluation step 203 is performed, if the performance of the learned classifier satisfies a detection rate (false positive rate), the learning is stopped and a procedure of determining the classifier as a final classifier is performed do. On the other hand, when the performance of the learned classifier does not satisfy the desired value, a new classifier is added to repeat the above-described learning process.

Here, when the unit classifier is added and the learning process is repeated, a procedure of reconstructing (204) the negative samples currently used is performed. At this time, a negative data set having the same size as that of the initially used negative sample (N) is constructed including a false positive test image even though the object is not a recognition target object. In addition, it is preferable that the added unit classifier uses a new type of feature information that has not been used before as an input value in the learning.

This learning process is repeated until the learning rate of the learning device is satisfied and the false positive rate is satisfied. In the case of the minimum accuracy, since the learning is not repeated, the reference value the threshold can be adjusted to achieve high accuracy. In addition, since the false positive rate can be filtered by the new unit classifier, the performance is improved every time the unit classifier is added.

3 is a diagram illustrating a procedure of a parallel learning method of a cascade classifier for object recognition according to an embodiment of the present invention.

Referring to FIG. 3, a parallel learning method of a cascade classifier for object recognition according to an embodiment of the present invention will be described in detail.

As shown in the drawing, the parallel learning method of the cascade classifier according to the present invention first carries out a learning preparation procedure. As described above with reference to FIG. 2, the learning preparation procedure includes a procedure of preparing and receiving a learning image or learning image data to be used for learning, a procedure of loading an image or an image from the data, And a process of extracting features or descriptor information (S10).

When the data to be used for learning is prepared, a procedure for inputting a sample to a plurality of unit classifiers included in the cascade classifier is performed. The sample input to the plurality of unit sorters includes a negative sample (N ') used in common with the positive sample (P) in the input data. At this time, the input negative sample N 'is previously prepared sample data that is commonly used and can be simultaneously input to a plurality of unit classifiers since the negative sample dependency is not involved in cascade classifier learning. Therefore, since the samples are input to the plurality of unit classifiers at the same time, it is possible to implement parallel processing capable of simultaneously processing a plurality of classifiers (S20).

Here, the commonly used negative sample N 'is a concept that is not used in learning through a conventional cascade classifier. In order to generate a parameter for discriminating between true and false in learning, positive data and negative data Since the learning is not performed by providing only the positive data, a common negative sample N 'prepared in advance is used.

Next, the unit classifiers perform unit learning through the input samples, and proceed to a procedure of evaluating each learning result (S22). Further, in this step, all the same common negative samples N 'are used, so that the learning proceeds mainly on the positive sample P without a negative sample reconstruction procedure, so that it is referred to as a positive learning step. The learning process using the normal negative sample N included in the input data is performed separately later, and the procedure is referred to as a negative learning step S32. The negative learning step will be described later.

On the other hand, in the positive learning execution and evaluation step, learning is performed while focusing on a detection rate, and in the negative learning step S32 described later, learning is performed with a focus on lowering a false positive rate do.

In the positive learning execution and evaluation step (S22), each of the plurality of unit classifiers performs a post-learning evaluation procedure. In the positive learning execution and evaluation step (S22), a procedure is performed to evaluate the performance and speed as a result of learning by each unit classifier do.

After the results of the evaluation are determined by the plurality of unit classifiers, a procedure for selecting a priority order for a plurality of unit classifiers based on the determination result is performed (S24). As the basis for selecting the priority, information such as detection rate, false positive rate, and time required of the individual unit classifier is used.

In this case, the priority to be determined is the order in which the recognition target image is input in the cascade classifier among the plurality of unit classifiers. That is, order information for a plurality of unit classifiers sequentially connected in the cascade classifier. In this way, when the priority of the unit classifier is determined, a predetermined sequence is generated. Therefore, when there are not many devices to be used in the recognition and analysis processing, a classifier can be generated even with a small calculation.

However, in the present invention, it is needless to say that the connection order of the unit classifiers can be determined in various ways without determining the connection order of the unit classifiers according to the above-described priority selection process. In other words, since the detection rate may decrease as more unit classifiers are connected, the unit sorter that records the highest detection rate among the unit classifiers that have reached the desired false positive rate may be selected, or more And a method of selecting a unit classifier in which classification processing is performed within a short time.

After the learning and evaluation procedures are performed in the plurality of unit classifiers, a negative sample N among the training data prepared in step S10 is input to the unit classifier (S30). At this time, only the negative sample (N) is used as the input value without the positive sample (P), and at least one negative sample (N) of the feature information extracted from the learning data other than the common negative sample .

In addition, when the negative samples N are input to the unit sorter, the input processing can be performed in the order of the unit sorters according to the priority order selected in step S30.

When the negative classifier is input to the unit classifier, the unit classifier performs learning using the input negative sample N and proceeds with the evaluation procedure (S32). At this time, since only the negative sample N is used in performing the learning but the detection rate and the false positive rate must be calculated in the evaluation procedure, the positive sample P can be used have.

Further, learning using only the negative sample N is referred to as a negative learning performing and evaluating step S32 as described above, and it is desirable to lower the false positive rate in such negative learning . That is, since such negative learning is often performed when an object other than a target object is input when performing object recognition in an image, it is more meaningful to lower a false positive rate.

In this manner, unit learning is performed while unit classifiers are connected one by one in step S32. If the target false positive rate is not reached according to the evaluation result after the evaluation procedure (S34) A procedure for repeating the negative learning is performed again by adding a classifier.

At this time, the negative sample N used for learning and evaluation is used again in a dependent manner. After performing the process of reconstructing the used negative sample N (S36), the reconstructed negative sample is added to the newly added unit Proceed with the procedure to input to the classifier.

Here, the negative sample reconstruction procedure S36 adds new negative samples that have not been used for learning in order to use negative sets determined to be false positives depending on the next unit classifier, This is the procedure for reconstructing the negative learning set to the same size as the negative set it did.

On the other hand, when the target result reaches the evaluation result (S34) in the step S32, the unit classifier in which the learning and evaluation is performed is determined as the final unit classifier, and a procedure for completing the cascade learning procedure is performed (S38).

4 is a diagram illustrating a process of performing a parallel learning method of a cascade classifier for object recognition according to an embodiment of the present invention.

Referring to FIG. 4, a parallel learning method of a cascade classifier for object recognition according to an embodiment of the present invention can be divided into a positive learning process (304) and a negative learning process (305). The positive learning process 304 includes a process in which a plurality of unit learning 301 is performed by inputting a positive sample P and a common negative sample N 'to a plurality of unit classifiers, And an evaluation process 302 is performed. After the learning and evaluation are respectively performed through the plurality of unit classifiers, the priority classifying unit 303 is performed based on the evaluation information.

As shown in the figure, after the positive learning 304 process is performed, a negative learning process 305 is performed. In the negative learning step 305, the unit learning 306 is performed by inputting only the negative samples N to the unit classifier, and evaluation 307 is performed through the unit learning 306. In the negative learning step (305), a new unit classifier is added until the result of the evaluation (307) satisfies the determined result, and iteratively proceeds. Further, in the iterative learning, a negative sample reconstruction (308) procedure is performed on the negative sample (N) to use the negative sample (N) used in the previous learning process.

When the negative sample is reconstructed in this manner, the negative learning 305 is repeated with the corresponding negative sample (N) as an input value to the next unit classifier.

On the other hand, if the evaluation result (307) satisfies the determined result, the unit classifier up to this time is determined as the final unit classifier, and the parallel learning procedure for the cascade classifier is completed.

As described above, the parallel learning method of the cascade classifier for object recognition according to the present invention during the learning of the cascade classifier divides the positive learning 304 and the negative learning 305, A plurality of unit classifiers are processed in parallel in the process, parallel processing can be implemented at the time of learning, and the learning process can be performed more quickly and efficiently.

The parallel learning method of the cascade classifier for sifting according to an embodiment of the present invention described above can also be implemented in the form of a recording medium including a computer program stored in a medium executed by a computer or a command executable by a computer . Computer readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. In addition, the computer-readable medium may include both computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Communication media typically includes any information delivery media, including computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave, or other transport mechanism.

While the method and system of the present invention have been described with reference to particular embodiments, some or all of their components or operations may be implemented using a computer system having a general purpose hardware architecture.

While the present invention has been described in connection with what is presently considered to be practical exemplary embodiments, it is to be understood that the invention is not limited to the disclosed embodiments, but, on the contrary, It is to be understood that the invention may be embodied in other specific forms. It is therefore to be understood that the above-described embodiments are illustrative in all aspects and not restrictive. The scope of the present invention is defined by the appended claims rather than the detailed description, and all changes or modifications derived from the scope of the claims and their equivalents should be construed as being included within the scope of the present invention.

101: unit classifier 102: object not the target object
103: the object

Claims (1)

A learning processing method in a cascade classifier used for object recognition processing on an image or an image,
A learning preparation step of extracting feature information on a learning image to be input as learning data to a plurality of unit classifiers;
A negative sample used in common with at least one positive sample of the extracted feature information is input to a plurality of unit classifiers, learning is performed, the evaluation is performed to determine the result, and then the plurality A positive learning step of determining the priority of the unit classifier of And
At least one negative sample of the extracted feature information is sequentially inputted to the unit classifier according to the determined priority order to perform learning, and when the desired result is obtained by the evaluation, the final unit classifier is determined, And a negative learning step of reconstructing a negative sample to be input next to a negative sample to be input next, and repeating the learning, in a unit sorter of the next step if not.

Images