Ψ-NET: A Novel Encoder-Decoder Architecture for Animal Segmentation

This paper proposes a novel Ψ-Net architecture that consists of three encoders and a decoder for animal image segmentation. The main characteristic of our proposed architecture is that the outputs at each depth level of the three encoders are summed up and then concatenated in the corresponding depth levels of the decoder for the upsampling process. We col-lected a new dataset consisting of 200 images for training the model, and we manually labelled the ground truth segmentation masks for these images. We trained our proposed model Ψ-Net on this dataset and compared the segmentation accu-racy with the classical U-Net and Y-Net architectures. Our proposed model achieved the highest accuracy on the dataset with 93% pixel accuracy, and 81.6% mean intersection-over-union (IoU).