Instance segmentation of group-housed pigs based on recurrent residual attention

Objective  To realize high-precision segmentation of individual pigs under different conditions such as pig adhesion and debris shielding in a group breeding environment.

Method  A total of 45 group-housed pigs of 20 to 105 days from eight sheds in real farming scenes were studied. Mobile camera images were used as data sources, and data enhancement operations such as changing brightness and adding Gaussian noise were performed to obtain 3 834 annotated pictures. We explored multiple models with the cross-combinations of two backbone networks ResNet50, ResNet101 and two mission networks Mask R-CNN, Cascade mask R-CNN. We also introduced the idea of recurrent residual attention (RRA) into the two major task network models to improve the feature extraction ability and segmentation accuracy of the model without significantly increasing the amount of calculation.

Result  Compared with Cascade mask R-CNN-ResNet50, Mask R-CNN-ResNet50 improved AP_0.5, AP_0.75, AP_0.5-0.95 and AP_{0.5-0.95-large} by 4.3%, 3.5%, 2.2% and 2.2% respectively. Different numbers of RRA modules were added to explore the impact on the prediction performance of each task model. The experiment showed that adding two RRA modules had the most obvious improvement effect on each task model.

Conclusion  The Mask R-CNN-ResNet50 model with two RRA modules can more accurately and effectively segment group-housed pigs under different scenes. The results can provide a model support for subsequent identification and behavior analysis of live pigs.