Deep learning-based defect detection on livestock operations

Abstract

The recent growth of computer vision system research has contributed to the advancement of livestock operations including meat processing. However, the defect detection of livestock products has not been actively studied due to the scarcity of dataset despite its significance. In the real world, the ratio of the defects in the livestock dataset is extremely low, compared to the normal products, resulting in insufficient model training. To address this problem, this study propose a deep learning-based anomaly detection method for an extremely unbalanced dataset. Adopting an anomaly detection framework from prior research, we suggest an adversarial autoencoder, which includes loss inversion to optimize reconstruction error for training skewed data, and a linear perturbation method called the Fast Gradient Sign Method (FGSM) to generate noises as anomaly cases. The model was then evaluated on the real-world dataset acquired from a livestock factory to detect defective chicken carcass (e.g., broken leg, injured skin, bent wing). The experiment results show that the proposed model outperforms existing models used for anomaly detection problems (i.e. DevNet and DeepSAD) and loss inversion and FGSM successfully improve the model performance when detecting livestock defects, even if the dataset contains extremely few defects.