Automatic counting of chickens in confined area using the LCFCN algorithm

Abstract

Grouping chickens based on their weights is an important process that takes place in many chicken farms in New Zealand where chickens are grouped into three categories: small, medium and large. Each category has pins (cages) to temporarily hold the chickens during the process and a permeant bigger section to hold the chickens after grouping. Chickens are weighed and placed in respective pins. Thereafter they are released to the permanent section. Currently, the chickens are counted manually when they are released from a pin to a bigger section. The task of weighing chickens, placing them in a pin and releasing them to a bigger section is repeated until all chickens are moved to their respective bigger section and the total number of chickens in each section is calculated. This manual effort is done by several employees and takes several hours. This study investigated the feasibility of using deep learning algorithms to replace the manual counting. We applied the localized fully convolutional network (LCFCN) algorithm to count and locate chickens from images of the pins. LCFCN was applied to a dataset of 4092 images containing 114132 chickens. The algorithm was evaluated using the Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) metrics and achieved the values of 0.5592, 1.36% and 1.67 respectively which are promising results in this setting. Furthermore, we modified the implementation of LCFCN to enable a user to manually alter the predicted labels to guarantee error free counting and localization.