Construction of Vermin Database to Improve Image Recognition for Protection of Cultural Heritages

Abstract

Raccoons are invasive aliens in Japan and damage ecosystem, crops, and buildings, especially, cultural heritages made of wood. We investigate vermin recognition for intelligent surveillance cameras to protect cultural heritages and we evaluated image recognition using convolutional neural networks (CNNs) with images we took in a wildlife breeding facility as training and test image sets. In this paper, raccoon images actually invading temples or shrines taken by surveillance cameras are used for evaluation. Three sets of images are used for training CNN: a set of raccoon images from a general training image set named CIFAR-100, a set taken in the facility, and a combination of the both. While CNN models trained with images from CIFAR-100 and from