Channel Randomisation with Domain Control for Effective Representation Learning of Visual Anomalies in Strawberries

Taeyeong Choi, Grzegorz Cielniak

Keywords: Agricultural Robotics, Precision Agriculture, Computer Vision, Anomaly Detection, One-class Classification, Self-supervised Learning, Representation Learning, Data Augmentation

TL;DR: We present novel strategies for effective application of "Channel Randomisation" to perform self-supervised learning of visual anomalies in strawberries

Abstract: Channel Randomisation (CH-Rand) has appeared as a key data augmentation technique for anomaly detection on fruit images because neural networks can learn useful representations of colour irregularity whilst classifying the samples from the augmented “domain”. Our previous study has revealed its success with significantly more reliable performance than other state-of-the-art methods, largely specialised for identifying structural implausibility on non-agricultural objects (e.g., screws). In this paper, we further enhance CHRand with additional guidance to generate more informative data for representation learning of anomalies in fruits as most of its fundamental designs are still maintained. To be specific, we first control the “colour space” on which CHRand is executed to investigate whether a particular model— e.g., HSV , Y CbCr, or L*a*b* —can better help synthesise realistic anomalies than the RGB, suggested in the original design. In addition, we develop a learning “curriculum” in which CH-Rand shifts its augmented domain to gradually increase the difficulty of the examples for neural networks to classify. To the best of our best knowledge, we are the first to connect the concept of curriculum to self-supervised representation learning for anomaly detection. Lastly, we perform evaluations with the Riseholme-2021 dataset, which contains > 3:5K real strawberry images at various growth levels along with anomalous examples. Our experimental results show that the trained models with the proposed strategies can achieve over 16% higher scores of AUC-PR with more than three times less variability than the naive CH-Rand whilst using the same deep networks and data.