Semantic Segmentation using Generative Adversarial Networks with a Feature Reconstruction Loss

The image-to-semantic labels classification is a very challenging task in image processing. Convolutional neural networks (CNN) have managed to achieve the state-of-the-art quality of the segmented image in semantic segmentation tasks. Still, the classification capability of such algorithms is not satisfactory to segment images that contain complex object boundaries and minimal regions. Recently, the Generative Adversarial Networks (GAN) were introduced, which can solve the overfitting of the generator network using the adversarial loss. In this paper, a GAN-based segmentation model is proposed, in which the Conditional GAN (CGAN) model is used as base architecture. Feature reconstruction loss is introduced in this composite model to solve the identification and classification of visually small elements in images. A pre-trained deep convolution neural network is adopted to calculate feature reconstruction loss. The usage of Feature reconstruction loss has demonstrated the high quality of the output labels and improved segmentation masks. The evaluation of the proposed model on the cityscapes dataset has shown the effectiveness of GAN-based architecture in semantic segmentation of multiclass images. The proposed model achieved 83.3% accuracy on the cityscapes test dataset, which is superior to most semantic segmentation state-of-the-art methods.