A Deep Learning Approach for Image Analysis and Reading Body Weight From Digital Scales in Pigs Farms

Abstract

Body weight is an important measure in fattening farms that allows pig farmers to monitor weight gain, manage feeding, and care for animal health. Traditionally, pig weighing is done directly, i.e. placing one or more pigs at a time on a scale while the total body weight is displayed and recorded. However, this process is labor-intensive, causes stress to the animals, and is highly prone to manual recording errors. Recently, several deep learning-based image analysis methods have emerged to estimate pig body weight, but these only consider the animal’s body characteristics. For this reason, an automatic deep learning-based approach is introduced for reading pig body weight from digital scale images. This reading is done by recognizing the body weight values indicated on the scale screen during the weighing process. For this purpose, convolutional neural network models are developed from scale screen segmentation to scale digit classification used to build the body weight value. The proposed approach is applied and validated in a real case study of fattening pigs from a Spanish company. Computational results showed that our approach read body weight with an average error of 20.2 g for a group of pigs with an average weight of 44 kg, taking less than 50 milliseconds to individually recognize the weight value. Therefore, our approach is reliable to support decisions in pig fattening management and suitable to be embedded into real-time weighing systems and useful too for image annotation purposes.

This work was supported in part by the Project ‘‘Smartfarms: Development of Digital Twin technologies in Pig Farms’’ under Grant 56 22 073 2021 2A; in part by the Project ‘‘AI4Pork: Artificial Intelligence in the Digital Transition of the Pig Sector’’ under Grant TED2021-130829B-100; in part by the Ciencia y Tecnología para el Desarrollo (CYTED) Network ‘‘AI for Agriculture in Iberoamerica’’ under Grant 524RT0158; and in part by the Chilean Agency of Research and Development (ANID) under Grant 75240020