Towards automatic wild animal monitoring: Identification of animal species in camera-trap images using very deep convolutional neural networks

Abstract

Abstract. Non intrusive monitoring of animals in the wild is possible using camera trapping framework, which uses cameras triggered by sensors to take a burst of images of animals in their habitat. However camera trapping framework produces a high volume of data (in the order on thousands or millions of images), which must be analyzed by a human expert. In this work, a method for animal species identification in the wild using very deep convolutional neural networks is presented. Multiple versions of the Snapshot Serengeti dataset were used in order to probe the ability of the method to cope with different challenges that camera-trap images demand. The method reached 88.9% of accuracy in Top-1 and 98.1% in Top-5 in the evaluation set using a residual network topology. Also, the results show that the proposed method outperforms previous approximations and proves that recognition in camera-trap images can be automated.

Identifying animal species in camera trap images using deep learning and citizen science

“…We surpassed the best results reported by Gomez Villa et al. () (roughly 58% as estimated from their plots) who also used the SS dataset but excluded rare species (modelling 26 species) and used transfer‐learning to train all of their models. However, contrary to our approach, they pretrained their models on the ImageNet dataset while we used transfer‐learning between camera trap projects.…”

“…To address these issues, automatic classification of camera trap images has been a focus of research in computer vision and machine learning. Recent advances in using techniques from deep learning have enabled researchers to improve automatic species identification significantly (Gomez Villa, Salazar, & Vargas, 2017;Norouzzadeh et al, 2018).…”

“…(), were not comparable to human accuracies of 96.6% (Swanson et al., ). More recent studies in automatically identifying animal species using CNNs have reported accuracies around 90% with some manual preprocessing (Gomez Villa et al., ) or more complex pipelines involving automatic preprocessing (Giraldo‐Zuluaga, Salazar, Gomez, & Diaz‐Pulido, ). The most recent advances by Norouzzadeh et al.…”

Software to facilitate and streamline camera trap data management: A review

“…Various approaches for similar classification programs exist for other purposes, such as facial recognition software, and it is surprising that image classification in camera trapping is still performed manually (Yu et al, 2013). A few authors have attempted to develop species recognition processes, with varying success McShea et al, 2016;Norouzzadeh et al, 2018;Villa, Salazar, & Vargas, 2017;Wang, 2014;Yu et al, 2013). Yu et al (2013), for instance, developed a mechanism that extracts the foreground (the animal) from the background, analyses the features of the object, and finally classifies the images by a linear support vector machine algorithm.…”

“…Wang (2014) reaches similar accuracies ranging between 77% and 87%, depending on the exact classification method and dataset, while He et al (2016) reports lower levels of accuracy (34% and 38%; depending on method used). Most recently, Villa et al (2017) and Norouzzadeh et al (2018) have reported up to 98.1% and 96.6% accuracy, respectively, however, this drops significantly when incorporating an unbalanced dataset including uncommon species. Indeed, a disadvantage is that for each species, a high number of "practice"…”