Multi-scale Autoencoders in Autoencoder for Semantic Image Segmentation

Semantic image segmentation is essential for scene understanding. Several state-of-the-art deep learning-based approaches achieved remarkable results by increasing the network depth to improve performance. Using this principle, we introduce a novel encoder-decoder network architecture for semantic image segmentation of outdoor scenes called SAsiANet. SAsiANet utilizes multi-scale cascaded autoencoders at the decoder section of an autoencoder to achieve high accuracy pixel-wise prediction and involves exploiting features across multiple scales when upsampling the output of the encoder to obtain better spatial and contextual information effectively. The proposed network architecture is trained using the cross-entropy loss function but without incorporating any class balancing technique to the loss function. Our experimental results on two challenging outdoor scenes: the CamVid urban scenes dataset and the Freiburg forest dataset demonstrate that SAsiANet provides an effective way of producing accurate segmentation maps since it achieved state-of-the-art results on the test set of both datasets, 72.40% mIoU, and 89.90% mIoU, respectively.