Adapting SVM Image Classifiers to Changes in Imaging Conditions Using Incremental SVM: An Application to Car Detection

摘要

In image classification problems, changes in imaging conditions such as lighting, camera position, etc. can strongly affect the performance of trained support vector machine (SVM) classifiers. For instance, SVMs trained using images obtained during daylight can perform poorly when used to classify images taken at night. In this paper, we investigate the use of incremental learning to efficiently adapt SVMs to classify the same class of images taken under different imaging conditions. A two-stage algorithm to adapt SVM classifiers was developed and applied to the car detection problem when imaging conditions changed such as changes in camera location and for the classification of car images obtained during day and night times. A significant improvement in the classification performance was achieved with re-trained SVMs as compared to that of the original SVMs without adaptation. incremental SVM, car detection, constraint training, incremental retraining, transfer learning.