Out-of-distribution algorithms for robust insect classification

Mojdeh Saadati, Shivani Chiranjeevi, Aditya Balu, Talukder Zaki Jubery, Asheesh K Singh, Soumik Sarkar, Arti Singh, Baskar Ganapathysubramanian

Keywords: Out-of-distribution detection, insect classification, iNaturalist, energy-based model, maximum softmax probability, Mahalanobis distance

TL;DR: OOD detection for robust insect classification

Abstract:

Plants are exposed to various useful and harmful insect pests during their growth cycle. Accurate identification of these pests is critical for deciding on a timely and appropriate mitigation strategy with significant economic and environmental implications. Recent progress in deep learning-based approaches has resulted in insects exhibiting good accuracy. However, deploying them in the wild is still problematic since input images that are wildly out of the distribution (e.g., non-insect images like vehicles, animals, or a blurred image of an insect or insect class that is not yet trained on) can still produce insect classification. To counter this, methods that ensure that a model abstains from making predictions are needed. To address this issue, we leverage the out-of-distribution detection concept that showed promising results in detecting out-of-distribution data in dermatology tasks (Roy et al., 2022).
In our work, we evaluate the performance of state-of-the-art out-of-distribution (OOD)
algorithms on insect detection classifiers. These algorithms represent a diversity of methods of approaching an OOD problem. Additionally, we focus on extrusive algorithms -- i.e., algorithms that wrap around a pre-trained classifier without the need for additional co-training. We choose three OOD detection algorithms: (i) Maximum Softmax Probability (MSP), commonly referred to as the baseline algorithms, (ii) Mahalanobis distance-based algorithm, which solves the problem using a generative classification approach; and (iii) Energy-Based Model OOD detection algorithm, which exhibits SOTA for OOD detection. We perform an extensive series of evaluations of these OOD algorithms across two performance axes: (a) how the accuracy of the classifier impacts OOD performance and (b) how the degree of out-of-domain impacts OOD performance. Our analysis shows OOD detection algorithms can significantly improve from abstaining classification across different settings of models’ structures and datasets. Thus, our OOD-robust classifier improves user trust in using the application for insect-pests classification.