Classification of Toxic Ornamental Plants for Domestic Animals Using CNN

Veterinary medicine emphasizes accidents caused by toxic plants with domestic animals as an extremely important topic, as the right diagnosis can be crucial for the affected animal. In this work, we propose the classification of toxic ornamental plants, according to nine different categories, using five widely-known CNN architectures, namely: DenseNet, ResNet, VGG16, VGG19 and Xception. The rationale behind it is that the automatic identification of these types of plant can be a useful tool to help in the prevention of those accidents. The authors have carefully curated a database to support the development of this work, collecting images available on the Pinterest website, and also performing some important data pre-processing. This database was also made available as a contribution of this work. Transfer learning was employed by taking advantage of feature learned from the ImageNet dataset. We also analyzed the heat maps generated by the Layer-wise Relevant Propagation method, which allowed to observe the individual behavior of the best and worst architectures. The best performance was achieved using DenseNet, with an accuracy of 97.67%. That model managed to generalize very well, even to deal with noisy images, which are frequent in photos of decorative environments.