Cattle face generation using an adversarial network with bounding boxes of facial subregions
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Juiz de Fora (UFJF)
DOI
Abstract
Data collection in precision livestock farming is a serious challenge. The scarcity
of annotated image datasets for precision livestock problems limits the potential of deep
learning algorithms. The information learned by the Generative Adversarial Networks
(GANs) can be employed to improve the performance of machine learning tasks, such
as classication. Specic subregions of the facial cattle region, such as the muzzle, ears,
and eyes, are signicant indicators of pain, discomfort, and stress in animals, as well as
being used for individual identication. In this context, in the present work we develop a
new architecture that enhances the spatial localization of facial interest subregions. The
proposed method combines bounding box modules with a GAN known in the literature,
named BigGAN. The hypothesis is that the incorporation of spatial location information,
provided by interest subregions bounding boxes, such as the muzzle, ears, and eyes,
can improve the generation of synthetic cattle faces images. Thus, we proposed a new
conguration of these modules, based on providing multiple objects in a scene. The
modules are attached at dierent depths in the generator, forming a resolution pyramid,
and at a specic level in the discriminator. We also propose to use a multiscale boxes to
represent the interest subregions. The premise is to smooth out the transitions between
the objects and the background, while reinforcing the objects’ features. To train the
model, we ltered a pre-annotated dataset for the detection task and adapted it for the
generation task. The new dataset includes 9, 495 cattle faces and a subset of bounding box
annotations. To assess the feasibility of the proposal, we conducted extensive analysis, both
quantitative and qualitative. The experimental results showed improvements provided
by the inclusion of additional information for the interest subregions via bounding box
modules. The capacity to arbitrarily insert the subregions’ position allows for possible
applications such as facial generation designed by user.
Description
Citation
Endorsement
Review
Supplemented By
Referenced By
Creative Commons license
Except where otherwise noted, this item's license is described as Acesso Aberto

