Identification of Poultry Reproductive Behavior Using Faster R-CNN with MobileNet V3 Architecture in Traditional Cage Environment

Abstract

Muscovy ducks are one of the waterfowl that are widely cultivated because of their relatively higher price compared to chickens and ducks. Determining whether Muscovy ducks are productive (raised to lay eggs and sell the eggs) or unproductive (raised/fattened to sell the meat) is one of the actions that can be taken to increase production. However, monitoring that only focuses on egg-laying behavior and is still done manually by humans is still prone to errors. An automated monitoring system is needed to improve monitoring. In this research, building a Deep Learning model using the Faster RCNN algorithm MobileNetV3 architecture to monitor poultry reproductive behavior. The identified behaviors are mating, non-mating, and egg-laying/swarming behaviors. Reproductive behavior can be used as a reference to determine whether a bird is productive or not. A model was built and tested on traditional cages. The addition of pre-processing techniques to improve image quality was performed. The results obtained were 86% mating behavior, 80% non-mating behavior, and 94% egg-laying. There was a difference in accuracy before using preprocessing techniques, which was 82% mating behavior, 74% non-mating behavior, and 88% egg laying. Adding Preprocessing methods (Image enhancement, ROI, and Blurring) can improve the performance of the Faster R-CNN algorithm to detect objects but impacts blurry image quality. The addition of deblurring techniques after the Faster R-CNN algorithm detection process can be used to restore image quality without affecting accuracy.