Object Detection with Dataset Augmentation for Fire Images Based on GAN

Abstract

Objection detection is the task to find and classify objects in images. Many object detection models based on a deep learning algorithm have been proposed. Deep learning algorithms require the models to be trained with affluent images with accurate annotations. However, in the case of fire detection, neither enough datasets to train the detection model nor correct and sufficient annotations exist. In this paper, we propose a GAN-based model to generate fire images with bounding boxes to enhance the performance of the fire detection model. Through the experiments, we demonstrated that the model can inject flame images into the clean images within specified areas and the generated images are enough to augment the fire detection dataset so that the model's object detection performance can be improved.