Faster R-CNN and EfficientNet for Accurate Insect Identification in a Relabeled Yellow Sticky Traps Dataset
Maurice Deserno, Alexia Briassouli
Abstract
Precision farming applications such as biological pest control use yellow sticky traps to catch insects and then classify and count them. A benchmark dataset of yellow sticky traps with insects has been recently made public for developing methods to identify and count them. However, this dataset contains missing or erroneous annotations, so we have corrected them and made them available <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> for reliable benchmarking. We use these corrected annotations to propose and compare approaches based on State-of-the-Art (SoA) object detection/recognition, namely Faster R-CNN and EfficientNet, with appropriate pre-processing and data augmentation. Pre-processing and augmentation is crucial for achieving accurate results with this dataset, which faces challenges caused by the small size of the objects, occlusions, similarity in appearance and other factors. Our experiments demonstrate that our architectures can lead to very accurate recognition of challenging insect classes, with accuracies from 81.27% to 99.1%, while future work is proposed for reducing false alarms and improving performance.