Semantic segmentation with deep learning and real image generation from pixel images

In the first part of this thesis, which consists of two parts, semantic segmentation is carried out with deep learning methods. Semantic segmentation is the process of associating each pixel in an image with a corresponding label. Semantic segmentation can be used to detect and locate objects in the image. Semantic segmentation has become an important issue for better interpretation and understanding of images by computer systems. In recent years, object detection from images with deep learning methods has been widely used in the interpretation of objects. Semantic segmentation was performed with the Deeplab v3+ CNN network based on the Resnet-18 transfer method. For this, the Camvid dataset was used. Pixel-based semantic segmentation was applied manually to the images in the dataset consisting of 701 high-resolution images. Primarily, the segmentation process was performed according to the Gretag–Macbeth color scheme. Then, Deeplab v3+ real images were matched with pixel images and the training process was carried out. Different images were used to test the model. High successes were observed according to the observed Jaccard, Sørensen-Dice and BF Score metrics. In the second stage of the thesis, synthetic images were created from pixel images with deep learning methods. For this, GAN methods, one of the deep learning methods, were used. GAN models are widely preferred to generate synthetic data in different fields. In our thesis study, Px2PxHD GAN model was used to create real images. Pix2PixHD is an image translation method used to produce realistic and detailed images from low-resolution maps of high-resolution images. The basis of this method is deep learning and especially convolutional neural networks. In Pix2PixHD GAN method, VGG19 transfer deep learning method was used as CNN network. Experiments were carried out on the Camvid dataset. It has been observed that successful high-resolution images were produced in the experiments carried out.