Modelo de aprendizado para classificação de escore de cocho em confinamentos de bovinos de corte com base em imagens

Abstract

The evaluation of feed bunk score is a fundamental technique for evaluating the leftover food in the feed bunk using notes or scores, which allows adjusting the amount of food supplied per day to optimize animal performance and reduce food waste. Therefore, the objective of this work was to develop a mobile technology capable of assisting in reading the feed bunk score in beef cattle feedlots by using digital images. To develop the platform, we initially included information on diet composition and feed bunk score images to create different bases according to the score. Concurrently, a study was conducted to identify the collected images based on chromatic aspects and computer vision. This information showed that in most feedlots sorghum, silage, barley, cottonseed, and cornmeal were used to compose the animals' diet. In addition, the most used bunks were cement and plastic, and H-bunk and J-bunk formats. Furthermore, the convolutional neural network proposed by this work showed that images in size 108 x 108 obtained higher accuracy (85.45%) in the classification by scores besides presenting a model size (59 M.B.) that would be possible embedded in an application. Thus, the model showed promising performance for both the score-based classification task and the one based on adjustments of the amount of food.