65 Using 3D images and deep learning to predict feeder occupancy in grow-finish pigs

Abstract A common source of outlying data in automatic feed recording systems used with grow-finish pigs is multiple pig occupancy of a single space feeder. Tracking these events is difficult and may require continuous observation or time-consuming manual decoding of video recordings. An alternative to manual decoding is use of algorithms for automatic image classification. The aim of this work was to build a hierarchical model to classify images from automatic feeders to discern if the feeder is empty, occupied by one pig, or occupied by more than one pig. An experimental pen was equipped with a ceiling-mounted Intel RealSense D435 camera above an automatic feeder. Images were recorded for 3 consecutive days. Randomly selected Images from d 1 were used for model training (n = 414 and n = 238 for models 1 and 2, respectively). Images from subsequent days were used to test the models (n = 196 and n = 427 for models 1 and 2, respectively). Images were manually labeled as “empty feeder,” “one pig inside the feeder,” or “more than one pig inside the feeder.” Pixel intensity data from depth images were used as predictors of feeder occupancy using a two-step decision tree that consisted of fitting two binary outcome models to a) predict if the feeder was occupied and b) predict if there was more than one pig in the occupied feeder. Both models consisted of a feed-forward artificial neural network. The transfer functions used were the reflected linear unit for deep layers and softmax for the outcome layer. The loss function was the binary cross entropy and the optimizer was the stochastic gradient descent. Accuracy, sensitivity, and specificity were 100% for model 1. For model 2, accuracy=97.7%, sensitivity=97.2% and specificity=98.1%. The proposed models allow accurate, continuous, real-time monitoring of feeder occupancy by no pig, one pig, or more than one pig.