Utilizando redes neurais convolucionais siamesas para filtragem de imagens vazias em dados de armadilhas fotográficas

Abstract

Camera trap images are used to non-invasive wildlife monitoring through tasks such as species population analysis and studying animal behavior throughout the seasons. However, since the images are obtained when the camera’s motion sensors are triggered, several images without animals are captured due to the fact that the motion sensors are triggered by other elements, such as trees and leaves. This results in an accumulation of empty images that use memory space and consume bandwidth and network energy. To solve this problem, it is necessary to use methods that allow filtering empty images. However, this a challenging task due to several characteristics, such as the variation of vegetation between different locations and throughout the day and the seasons. In this context, the objective of this work is to present an approach to filter empty images captured by camera trap devices which takes into account information of the environment surrounding the camera. The proposed approach is based on a Siamese convolutional neural network that works with two input images: 1) an image without animals that presents the characteristics of the local vegetation and the approximate level of daylight; and 2) an image captured as usual due to the motion sensor triggering, which will be checked to determine whether or not there are animals in the scene. When processing the two images, the Siamese network identifies the semantic differences between them so as to identify the presence of animals in the captured image. The obtained results indicate that the Siamese approach reached superior precision and accuracy rates when compared to models that deal with only one image at a time, such as canonical convolutional neural networks.