Developing and Evaluating Poultry Preening Behavior Detectors via Mask Region-Based Convolutional Neural Network

Abstract

There is a lack of precision tools for automated poultry preening monitoring. The objective of this study was to develop poultry preening behavior detectors using mask R-CNN. Thirty 38-week brown hens were kept in an experimental pen. A surveillance system was installed above the pen to record images for developing the behavior detectors. The results show that the mask R-CNN had 87.2 ± 1.0% MIOU, 85.1 ± 2.8% precision, 88.1 ± 3.1% recall, 95.8 ± 1.0% specificity, 94.2 ± 0.6% accuracy, 86.5 ± 1.3% F1 score, 84.3 ± 2.8% average precision and 380.1 ± 13.6 ms·image−1 processing speed. The six ResNets (ResNet18-ResNet1000) had disadvantages and advantages in different aspects of detection performance. Training parts of the complex network and transferring some pre-trained weights from the detectors pre-trained in other datasets can save training time but did not compromise detection performance and various datasets can result in different transfer learning efficiencies. Resizing and padding input images to different sizes did not affect detection performance of the detectors. The detectors performed similarly within 100–500 region proposals. Temporal and spatial preening behaviors of individual hens were characterized using the trained detector. In sum, the mask R-CNN preening behavior detector could be a useful tool to automatically identify preening behaviors of individual hens in group settings.

Practices and Applications of Convolutional Neural Network-Based Computer Vision Systems in Animal Farming: A Review

“…Most of the applications transferred CNN architectures pretrained from other publicly-available datasets. RGB images were built and annotated in these datasets, and using RGB images may improve the efficiency of transfer learning [ 51 ]. Including depth information may further improve the detection performance.…”

“…General solutions are to manually remove invalid files and retain informative data for the development. Examples of invalid files include blurred images, images without targets or only with parts of targets, and images without diverse changes [ 11 , 51 , 80 , 90 , 107 ]. Some image processing algorithms were available to compare the differences between adjacent frames or between background frames and frames to be tested and capable of automatically ruling out unnecessary files.…”

“…Tools for semantic/instance segmentation can assign attributes to each pixel. These included Graphic [ 92 ], Supervisely [ 104 ], LabelMe [ 64 , 90 , 112 ], and VGG Image Annotator [ 46 , 51 ]. Labels of pose estimation generally consist of a series of key points, and available tools for pose estimation were DeepPoseKit [ 97 ] and DeepLabCut [ 132 ].…”

“…This strategy can save training time for complex models but does not compromise detection performance. Various similarities between datasets of transfer learning and customized datasets can result in different transfer learning efficiencies [ 51 ]. Figure 10 shows popular public datasets for transfer learning in animal science, including PASCAL visual object class dataset (PASCAL VOC dataset), common objects in context (COCO) dataset; motion analysis and re-identification set (MARS), and action recognition dataset (UCF101).…”

“…Different loss functions are generally used to evaluate the errors between ground truth and prediction during training. These include losses for classification, detection, and segmentation [ 51 , 111 ]. Accuracy over different steps is also applied during training [ 126 ].…”