Deep learning based thermal image segmentation for laboratory animals tracking

Automated systems for behaviour classification of laboratory animals are an attractive alternative to manual scoring. However, the proper animals separation and tracking, especially when they are in close contact, is the bottleneck of the behaviour analysis systems. In this paper, we propose a method for the segmentation of thermal images of laboratory rats that are in close contact during social behaviour tests. For this, we are using thermal imaging - a technology that requires neither light nor human presence. The aim of the study was: (1) an efficiency analysis of deep learning based image segmentation algorithms for the need of laboratory rats images, (2) analysis of different methods of original thermal data conversion to grey scale images for the purpose of the segmentation, (3) evaluation of the image data range impact on segmentation results using deep learning networks. We have trained U-Net and V-Net architectures with images obtained from different temperature ranges. The results indicate, that networks trained on images containing a narrow range of temperature data equal to animals' body temperature or even its part, are able to better perform the object segmentation than networks trained on the original data.