Automatic detection of injuries in turkeys : dealing with the prerequisites for a consistent annotation

In fattening turkeys, injurious pecking is a major health and welfare issue. Abnormal pecking behaviour occurs particularly when bleeding injuries act as a trigger. Earliest detection of these injuries is required to start intervention to prevent larger outbreaks. Thus, the aim of the presented study was to develop a commercial system for automatic detection of injuries in turkeys. As a mandatory initial step in the system development, a neural network was trained. Here, first insights into the agreement between the network and human observers is presented. Fattening turkey hens were video-recorded, and the videos were sliced into individual images for further processing. In the first step, these images were added to a developed annotation software that enables humans to mark pecking injuries in the images. Based on these manually labelled data, a neural network was trained to detect and label injuries in newly recorded images. The agreement between the observers and the network was not acceptable (32%). Therefore, a second step was added, where the previously made annotations were verified by presenting them a second time to the human observers in a modified software. In this software, a pseudo-binary system was created, and the observer had to confirm a correctly classified detection (‘injury’ versus ‘no injury’) while being blinded to the initial decision originating from the first step. A total of 24,173 opinions were given by human observers with 18.5% labelled detections rated as ‘no injury’, while 81.5% were classified as ‘injury’. In 81.6% of all labelled injury detections, the observers agreed on the classification. With the creation of a pseudo-binary system, the agreement between the human observers was improved substantially. In a future step, the network will receive the unanimous results from this second step for further training to improve its performance.