Incremental Object Detection Using Ensemble Modeling and Deep Transfer Learning

Conference proceedings article

Authors/Editors

Strategic Research Themes

Publication Details

Author list: Piyapong Huayhongthong, Siriyakorn Rerk-u-suk, Songwit Booddee, Praisan Padungweang and Kittipong Warasup

Publisher: Springer

Publication year: 2020

Volume number: 1149 AISC

Start page: 190

End page: 198

Number of pages: 9

ISBN: 9783030440435

ISSN: 2194-5357

eISSN: 2194-5357

URL: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85083679620&doi=10.1007%2f978-3-030-44044-2_19&partnerID=40&md5=0d63bec780b1e33e5a3dbf59b8d0043d

Languages: English-Great Britain (EN-GB)

View on publisher site

Abstract

Object detection is a subset of computer vision that can be accomplished using machine learning. The main process of object detection using machine learning model is model training with images containing objects of interest. However, the model training need a lot of training images. In addition, to improve the model ability to detect addition class of object, it need to be re-trained with both old and new image datasets. It is a time and computation consuming process. This paper proposes an incremental object detection model without re-training the old images. An ensemble model and transfer learning approach are used. The proposed model consist of three parts, two object detection sub-models and a decision model, which are a pre-trained model, a transferred-model and an ensemble model respectively. To illustrate the proposed model, the trained YOLO algorithm training with eighty object categories, 330,000 total images, from COCO image dataset is selected as the pre-trained model. It also be used as an initial model to train the transferred-model using transfer learning technique. Only new images are used for transferred-model training. The ensemble model with the bagging technique is used as a final classifier for choosing the best decision from both sub-models. Using our proposed model, the need of both the number of training dataset and the training time are reduced. Only several hours are needed for model training with three new object categories, 3,000 total images. The experimental results show that the proposed model achieve high performance on test image dataset with 93.33% accuracy. © 2020, The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG.

Keywords

Common object in object (COCO), Deep transfer learning, Ensemble modeling, You Only Look Once (YOLO)