標題: 線上增強式學習偵測系統
On-line Reinforcement Learning-Detection System
作者: 鄭丁榮
林進燈
林建中
Zheng,Ding-Rong
Lin, Chin-Teng
Lin, Chien-chung
光電系統研究所
關鍵字: 線上學習;偵測;增強;On-line Learning;Detection;Reinforcement
公開日期: 2016
摘要: 本研究提出一種線上增強學習偵測方法可讓使用者在影片中隨意的選取自己想偵測之目標種類,透過選取少量影像樣本生成初始偵測模型,並隨著影片播放時能夠接續且自動訓練此偵測模型,並對於不同種類的物體皆有相當程度的偵測結果。
本研究之架構主要分為兩部分,初始化偵測模型與線上學習訓練部分。初始化偵測模型階段藉由使用者所框選的樣本長寬比例基於Haar-like特徵產生特徵池,再由框選之訓練樣本從特徵池中訓練產生分類器選出較有辨別力之分類器並生成正負樣本比對模型,完成偵測模型初始化。接著在影片播放時藉由偵測模型所偵測出的新樣本用NN Classifier與正負樣本比對模型算出新樣本的正負樣本相似度後,融合背景模型計算出新樣本和欲偵測目標的相對相似度,再由此相對相似度保守的分類新樣本,保守分類後的正負樣本隨著影片播放進行自動的線上學習訓練來持續的更新分類器、更新正負樣本比對模型和選取新的分類器以持續的更新偵測模型。
測試本研究用於不同種類物體之結果,皆能藉由少量的選取偵測目標樣本再加上自動的線上學習訓練,有效減少了人力收集分辨大量的影像樣本和訓練大量時間,而測試結果皆也得到了不錯的偵測能力。
This study proposes a method that allows users to freely select what target species they want to detect, generate initial detection model by selecting a small amount of image sample, and as the movie plays that can continue training this detection model automatically, and for different types of objects, our method has considerable degree of detection results.
Framework of this study is divided into two parts, the initialize detection model and online learning section. Detection model initialization phase by the sample size based on the proportion of users of the Haar-like features to generate a pool of features, and then by the training samples generated from the feature pool to train and select effective classifiers, complete the initial detection model. Then the movie plays by detecting model detects the new sample with NN Classifier with positive and negative samples of positive and negative samples than similarity model calculates new samples after fusion background model to calculate a new sample and want to detect the target relative similarity, and from this relative similarity conservative classification of new samples, positive and negative samples conserved classified with the video player automatically online learning and training to continuous updates classifier, positive and negative samples than updating the model and select a new classifier to continually update detection model.
In this study, the results of the test for different types of objects , able to detect the target by choosing a small number samples plus automatic online learning , effectively reducing the manpower to collect a large number of image Samples and a lot of time for training, and the test results are also obtained good detection capability
URI: http://etd.lib.nctu.edu.tw/cdrfb3/record/nctu/#GT070258011
http://hdl.handle.net/11536/143004
顯示於類別:畢業論文