具適應性特徵集合之線上多類別學習

Abstract

本研究提出一種新的線上多類別學習演算法，此演算法主要有三個特點。首先，在特徵萃取上提出具適應性特徵集合，藉由動態結合Haar-like、HOG與LBP三種通用特徵，使特徵集合能更適應樣本集合。接著是可整合外部模型，在原先模型學習完畢後，可以在不重新訓練原先模型的前提下去外掛外部模型，增強模型的效能。最後是偵測模型的部分，本研究提出一種新的多類別學習與更新機制，模型能自主發現不適合的決策並自動調整。 在本論文中，本研究驗證了模型在多類別偵測與線上學習的表現，並在公開的行人與多類別數據庫中，擁有比其它非深度學習演算法更良好的效能，其中多類別數據庫裡包含：行人、人臉、車輛、機車、腳踏車與飛機。

We propose a new online multi-class learning algorithm. There are three main characteristics in this algorithm. First, in order to make the feature pool fitter for the pattern pool, we propose adaptive feature pool to dynamically combine the general three features, Haar-like, HOG, and LBP, in our feature extraction. Besides, we can integrate the external model into our completed original model without re-training to enhance the efficacy of the model. Last but not least, we propose a new multi-class learning and updating mechanism, which can find the unsuitable decisions and adjust it automatically. In this thesis, we validate the performance of the proposed model in multi-class detection and online learning. We achieve a better score than other non-deep learning algorithms used in public pedestrian and multi-class databases. The multi-class databases contain pedestrians, faces, vehicles, motorcycles, bicycles, and aircraft.