10 Deep Learning for Multi-Behavioral Video Classification of Interactive Behaviors of Pigs in Single-Spaced Automatic Feeders

Abstract

Abstract Automatic pig feeders record individual feeding behavior traits that can be used in genetic selection. However, these feeders cannot measure how feeding traits are affected by behavioral interactions between animals. Moreover, recording behavioral interactions can improve the estimation of social genetic effects. This motivated our assessment of computer vision to classify 4 types of interactions at the feeder: head-to-body contact (HB) including head knocking, tail biting, pushing, gentle nosing and casual contact between head/ears of a pig with a feeding pig; levering (L) where the feeding pig was lifted from behind by another pig; mounting (M) where the feeding pig was mounted by another pig; and no-contact (NC) when a second pig entered the feeder without touching the feeding pig. Behavior at the feeder was filmed twice, 3 weeks apart, for 2 consecutive days each week in 6 groups of grow-finish pigs (10 per group) housed in pens equipped with FIRE feeders. Video segments that involved 2 pigs in the feeder were selected, and labeled as HB (n=10,114), L (n=925), M (n=1,242) and NC (n=3,398). Due to the sparse data available for training and the complex nature of the behavioral interactions, we utilized pretrained convolutional neural networks for automatic spatial feature extraction followed by long short-term memory for temporal feature extraction from sequences of frames. Focal loss, a loss function that assigns different weights to hard/easily misclassified examples to handle class imbalance problem, was used in this study. Accuracy, recall, and precision of behavior classification were obtained from 5 random cross-validations. The overall accuracy was 0.967±0.002. The average recalls for HB, L, M, and NC were 0.977±0.004, 0.866±0.047, 0.969±0.011, and 0.964±0.009 respectively, while average precisions were 0.976±0.002, 0.901±0.029, 0.956±0.018, and 0.963±0.012. The proposed algorithm accurately classified multiple interactive behaviors in an automated feeding stall from digital videos.