초록

용어보기논문에서 용어와 풀이말을 자동 추출한 결과로, 시범 서비스 중입니다. AI-Helper

본 논문에서는 빠른 서포트 벡터 분류를 위해 신경망을 사용하는 방법을 제안한다. 주어 진 학습 데이터를 통해 낮은 학습 오류를 가지는 다단계 신경망을 얻으면 출력층을 제외한 은닉층은 주어진 문제를 선형분리 가능하게 하는 특징 추출기로 간주할 수 있다. 많은 계산시간을 요하는 키널 맵 대신 이를 사용해서 빠른 서포트 벡터 분류를 가능하게 하였다.

본 논문에서는 빠른 서포트 벡터 분류를 위해 신경망을 사용하는 방법을 제안한다. 주어 진 학습 데이터를 통해 낮은 학습 오류를 가지는 다단계 신경망을 얻으면 출력층을 제외한 은닉층은 주어진 문제를 선형분리 가능하게 하는 특징 추출기로 간주할 수 있다. 많은 계산시간을 요하는 키널 맵 대신 이를 사용해서 빠른 서포트 벡터 분류를 가능하게 하였다.

AI 본문요약

엑셀 다운로드 AI-Helper

* AI 자동 식별 결과로 적합하지 않은 문장이 있을 수 있으니, 이용에 유의하시기 바랍니다.

• 연구 주제
• 연구 방법
• 연구 결과

문제 정의

• To gain an insight into the performance of SVMs, an SVM with the polynomial kernel of degree 2 (4-(x, y)=(x-y)1 2 *) was applied to the LP detection problem [4]. The objective is to classify each pixel in an image into plate class or non-plate class based on their local color texture properties (RGB values of surrounding square image patch (I Ixl 1)). The SVM was trained on approximately 20.

제안 방법

• The trained ANN is then prunned based on the OBS algorithm which evaluate the demage of the network based on the quadratic approximation of increase of training error and prune it to obtain the minimal demage. Here we give a brief review of OBS.
• For details of experimental setting, readers are referred to [4]. For training the ANN+SVM classifier, 20.000 training examples which were used in training the base SVM classifier (Sec. 2.1) were used: the ANN was firstly trained on random selection of 10, 000 patterns. Then, the linear SVM was trained on the outp니t of the ANN feature extractor on whole 20, 000 patterns (including 10, 000 patterns used to train the ANN).

대상 데이터

• The training set was initialized with 2,000 patterns sampled from a database of 100 vehicle images and is augmented by performing bootstrapping [5] on the same database. The trained SVM was composed of around 1, 700 SVs and showed 2.7% training error rate. For testing, another set of 10, 000 plate and non-plate patterns were collected from 350 vehicle images which are distinct from the images used in training.
• 7% training error rate. For testing, another set of 10, 000 plate and non-plate patterns were collected from 350 vehicle images which are distinct from the images used in training. The SVM achieved 4.
• The proposed method was tested using an LP image database of 450 images. For details of experimental setting, readers are referred to [4].

이론/모형

• The basic idea of OBS is to use the second order approximation of error surfaces. Suppose for a given configuration of weights w (cf.

성능/효과

• There are already several methods to reduce the runtime complexity of SVM classifiers, in이tiding reduced set method [고], cascade architecture of SVMs [6], etc. The preliminary experiment with reduce set method has shown that at least 400 SVs were required to retain a moderate redaction of generalization performance, 3 which is still not enough to realize a practical LP detection system. The application of cascade method is beyond the scope of this work.
• The problem of high run-time complexity of SVMs was approached by utilizing a regularized ANN as the feature extractor. In comparison with the standard nonlinear SVM, classification performance of the proposed method was only slightly worse while the run-time is significantly better. Accordingly, it can provide a moderate alternative to the standard kernel SVMs in real-time applications.

본문요약 정보가 도움이 되었나요? 예 아니오 제출

이 논문을 인용한 문헌

더보기