Multi-Encoder Convolution Block Attention Model for Binary Segmentation

Behavioural research in animals can be assisted significantly with an automatic identification system. Such systems can evaluate animals’ behaviour non-intrusively, hence preserving their typical habitat. Recently, methods based on deep learning have shown promising results in this domain. In particular, object and key-point detectors have been used to detect individual animals. Although good results are obtained, bounding boxes and dispersed key points do not follow the animal’s contour, resulting in a large amount of information loss. This work proposed a binary segmentation model that precisely segments individual animal pixels in an indoor setting. In a nutshell, we proposed a new model with multiple encoders and a single decoder incorporating the attention mechanism. The method is tested on a specially created dataset with 1280 hand-labelled images and achieves detection rates of about 91% (dice coefficient) despite perturbations such as occlusions and illumination variations. The results are compared with state-or-the-art segmentation models, and a substantial boost in performance is achieved.