KR20150124825A - 2-Dependence Naive Bayes Classifier based Image Classification - Google Patents

Abstract

According to the present invention, 2-dependence naive bayes (2-DNB) is designed based on solving a vulnerable assumption of conditional independence of a feature in an image in a simple naive bayes nearest neighbor (NBNN) image classifier. Unlike other image classification methods (spatial pyramid matching (SPM), bag of words (BOG)), the 2- DNB takes an advantage of the NBNN, so the same is very simple and does not need a further learning/educational stage. Also, the 2-DNB is a unique non-parameter classifier designed based on a unique image space since the feature extracted from one image of earth verification data is dependent more than a research in other fields or related thereto.

Description

2-Dependence Naive Bayes Classifier based Image Classification < RTI ID = 0.0 >

The following embodiments relate to a simple Bayes Nearest Neighbor image classifier.

Since 1998, Naive Bay has been widely used as the basis for text classification because it can be implemented quickly and easily. It is difficult to apply the Naïve Bayes classifier to image classification, unlike text classification, where words can be matched exactly. However, in 2008, Boyman et al. Proposed a new, efficient, and nonparametric method for image classification, NaBNN (Naive Bayes Nearest Neighbor) classification.

Embodiments can provide a technique for solving the problem of conditionally independent assumptions when performing image classification in a simple Bay Nearest Neighbor.

According to one embodiment, a method of classifying images based on a 2-dependent simple Bayesian image classification may be provided.

According to one embodiment, the problem of conditionally independent assumptions can be solved when performing image classification in a simple Bay Nearest Neighbor.

FIG. 1 is a diagram for explaining a dense SIFT (Scale Invariant Feature Transform) feature.
2 is a diagram for explaining an algorithm of the NBNN.
FIG. 3 is a view for explaining proximal points in the feature space. FIG.
4 is a diagram for explaining an algorithm of 2DNB according to an embodiment.

It is to be understood that the specific structural or functional descriptions of embodiments of the present invention disclosed herein are presented for the purpose of describing embodiments only in accordance with the concepts of the present invention, May be embodied in various forms and are not limited to the embodiments described herein.

Embodiments in accordance with the concepts of the present invention are capable of various modifications and may take various forms, so that the embodiments are illustrated in the drawings and described in detail herein. It should be understood, however, that it is not intended to limit the embodiments according to the concepts of the present invention to the particular forms disclosed, but includes all modifications, equivalents, or alternatives falling within the spirit and scope of the invention.

Image classification is an item of pattern recognition in computer vision which is a classification method based on content information of an image.

In computer vision and image processing, a feature is a piece of information related to solving a computer task associated with a particular application. In particular, a feature can be the result of a common neighborhood task (feature extractor or feature detector) applied to an image, or a specific structure in the image itself, from a simple structure point or edge to a more complex structure, such as an object.

Conditional independence refers to the probability that another event will occur under the assumption of certain conditions. In the probability theory, if the occurrence or non-occurrence of B and the occurrence or non-occurrence of B are independent in the conditional probability distribution of Y given that the given third event Y is positive, then the two events R and B are conditional independent.

In probability theory, conditional dependence refers to the relationship between multiple events that depend on the occurrence of the third event.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings.

Queries are used interchangeably with inquiries. It refers to a query based on some code or key to know the contents of the file. It supports the function of retrieving the data existing in the database through the condition that the user wants and retrieving the retrieved result freely.

A descriptor is a term used to define the properties of a program unit in a computer or to define the characteristics of each record in each block of a file, and it is translated into "descriptor" and "descriptor". This descriptor is a short word that properly indicates the content of the information. It is used as a keyword to find the objective record and the target program in the file, and it is used for the system check that limits the use of the vocabulary. In the information retrieval, the name and the type name used to classify the information in the database or to add an index to the information, and the name and the type name are searched by the keyword and the purpose is investigated.

중 하나를 의미한다. NBNN에서, 디스크립터들은 주어진 그의 클래스 C에 독립적인 것으로 간주된다. 이는 양자화 단계를 생략함으로써 효율적이고 단순할 뿐만 아니라 이미지 분류를 개선하였다. 알려진 바와 같이, 양자화 단계는 감독 분류 방법에 필수적인 이미지의 일부 고유한 피쳐를 제거한다. 예를 들어, Bag of Words(BoW), ImageNet 등이 있다.Here Q is a query image, and I want to find its class C. Di is the descriptor of the query image Q

≪ / RTI > In NBNN, descriptors are considered to be independent of their class C given. This improved efficiency and simplicity as well as image classification by omitting the quantization step. As is known, the quantization step removes some unique features of the image that are essential to the supervised classification method. For example, Bag of Words (BoW) and ImageNet.

However, NBNN still has the following weaknesses: 1) the computational complexity is high during testing; 2) The original NBNN only works well on balanced data sets; 3) The assumption of independence of each descriptor violates the performance of the NBNN classification; 4) Although NBNN performs remarkably in image classification. Several optimization methods (optimized NBNN, class-to-image distance, NBNN kernel, local NBNN) have been proposed to solve the aforementioned problems in performing image classification for large image retrieval. However, no work has been done to address the problem of conditionally independent families that are most harmful to NBNN.

FIG. 1 is a diagram for explaining a dense SIFT (Scale Invariant Feature Transform) feature.

Referring to FIG. 1, a feature is dependent or related to lead to some concept or information. In particular, given a particular class of images, the distinctive features extracted from the images that make up an object or part of an object are much more dependent. Thus, in image classification, the image features are much more condition dependent and the assumption of conditional dependence of image features can be much more violent to classification. In addition, we can see that the proximal key points extracted from the image are most dependent or correlated.

Thus, instead of removing conditional independent assumptions, we use the dependence of the image's proximal features to solve the conditionally independent assumptions of the Naive Bay classification.

As mentioned earlier, some work has been done in the Naive Beit text classification to address the drawbacks of conditional independence. In a paper by Harry et al., We used a Bayesian network called the Augmented Naive Bayes classification (ANB) to increase the accuracy of classification results to calculate the likelihood of locally dependent features.

To obtain partial dependencies between attributes, you need to know the attribute's father point in the local bayesian network. However, unlike text classification, attributes in an image can not be as accurate as text matching. Therefore, if you want to model an augmented Naive Bayesian network (ANB) for image classification, you must find a difficult and time-consuming father feature. The cancellation chapter is exactly the same as the basic ANB, even if they have strong dependencies between properties, if the distribution of dependencies among properties satisfies certain conditions. With 2-dependence between proximal features, we designed the image classifier 2-DNB.

Naive Bay Near Neighbors

에 속하는 것으로 분류된다.To motivate and justify 2-DNB, we provide an overview of the original induction. In conjunction with the foregoing, the query image Q is classified according to Equation (2)

. ≪ / RTI >

Assume uniformity before all classes and apply Bayes rules.

의 조건 독립의 가정에 기초하여, 수학식 3은 수학식 4로 변환될 수 있다.The features extracted from the query image Q (descriptors)

(3) can be transformed into equation (4) based on the assumption of conditional independence of < RTI ID = 0.0 >

By means of a parzen window estimator, the kernel K, p is given by Equation (5).

In statistics, kernel density estimation (or Parzen window method, named after Emanuel Parzen) is a non-parametric method of estimating the probability density function (pdf) of a random variable. For example, given some data for a population sample, the kernel density estimation allows for an overall population estimate by extrapolation.

는 클래스 C에서 j 번째로 가까운 디스크립터이다. NBNN에서, j는 가장 가까운 이웃

를 사용하여 극단적이 된다.There are L descriptors in the training set for class C,

Is the closest descriptor in class C. In NBNN, j is the closest neighbor

It becomes extreme using.

K is selected as a Gaussian kernel, and Equation (6) can be transformed into Equation (7).

Therefore, NBNN can be arranged as shown in FIG.

The Bayesian classifier takes a parametric approach to estimate the probability distribution function of the data and apply it to classification. However, it is difficult to expect good performance if the pre-assumed probability model is not suitable for a given data distribution. An alternative to this problem is to use a K-neighborhood neighbor classifier based on non-parametric density estimates.

The K-neighborhood neighbor classifier finds K data in the order of closest distance from the given data, and assigns it to the class to which the largest number of data belongs. This classifier has a very nonlinear decision boundary and is based on a nonparametric method as compared with the regular Bayes classifier. Therefore, performance is not greatly influenced by the distribution of data.

In the case of the K-nearest neighbors classifier, it takes a long time to calculate the K neighbor data by calculating the distance to the entire learning data set every time new data is given, and there is always a problem that the learning data must be stored . In the case of K-proximity neighbor classifier, how to determine K value is a problem. If K = 1, noise is sensitive because it depends on only the closest one data. On the contrary, if the value of K is too large, the classification is performed based on the ratio of each class in the entire data area, not on the basis of only the area around the given data. Usually, K = sqrt (N) can be used when the total number of data is N, but it is also possible to think about finding the best performance value by performing classification using verification data.

According to one embodiment, assuming that there is a dependency of the proximal feature, we can pair the key point to the feature space.

Referring to FIG. 3, the point pair (A-B) defines dependency and the pair is conditional independent. Unlike other field studies, features in an image contribute to some concepts or information; In particular, it provides a single class to which it belongs. Therefore, the 2-dependent bayesian is defined as Equation (8).

과 함께 주어진 하나의 쿼리 이미지Q 이전의 고정형태는 수학식 9와 같이 변형될 수 있다.Therefore,

Can be transformed as shown in Equation (9). &Quot; (9) "

또는

및

일 때, 수학식 10은 하기 수학식 11과 같이 되지만, 2-DNB에서 k 값을 통일하지 않는다.According to 2-Dependent Bayesian

or

And

, Equation (10) becomes Equation (11), but does not unify the value of k in 2-DNB.

의 가장 가까운 이웃

을 찾아야하고,

가 속하는 모든 클래스 사이에서 디스크립터

의 최근접 이웃

을 찾을 것이다. 따라서, 2-DNB의 알고리즘은 도 4와 같이 나타날 수 있다.With a series of transformations, we can write descriptors

Nearest neighbor

However,

Between all the classes to which the descriptor belongs

Nearest neighbor

. Therefore, the algorithm of 2-DNB can be represented as shown in FIG.

So far, no work has been done to solve the weak assumption of conditional independence in the feature space of an image. As we study the features of image space, we have found that the descriptors extracted from one image act together to form an object or a solid part. In particular, for a given class belonging to one image, the feature classification of this class is more dependent. Thus, conditional independent assumptions can make more violations in image space.

According to the inspiration mentioned above, we define 2DN (2-Dependence Naive Bayes) using dependencies between proximal features. And we simply build a 2-DNB classifier to define a 2-dependent bayesian that works uniquely in feature space.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. For example, it is to be understood that the techniques described may be performed in a different order than the described methods, and / or that components of the described systems, structures, devices, circuits, Lt; / RTI > or equivalents, even if it is replaced or replaced.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (1)

In the image classification method,
2 Dependent simple Bayesian image classification based on image classification
Image classification method.

Images