Conditional Generative Adversarial Network(cGAN) for Image Super Resolution using class Embeddings

Abstract

The thesis titled ”Conditional GAN for Image Super-resolution using class Em beddings ” presents method based on conditional GAN architecture for addressing the challenging task of image superresolution. Image superresolution is a crucial research topic in computer vision that aims to obtain higher spatial resolution images from lower resolution inputs, without modifying camera hardware, making it a cost-effective and practical solution. The method incorporates class labels in the form of embedding layers which is used to provide conditioning to discriminator network. The method leverage the power of con ditional GANs to generate high-quality, visually consistent high-resolution images from lower-resolution inputs. To evaluate the performance of the model, generally metrics used for super-resolution are SSIM, MSE, and PSNR. Additionally here, a classifier trained on MNIST images is used to assess the quality of the generated images from both the conditional GAN (cGAN) model using embedding layer for labels and normal vanilla GAN model using one-hot encoded labels. The accuracy of 0.84 is obtained on generated image using cGAN with labels encoded with embedding layer and 0.69 on normal vanilla GAN with one-hot encoded labels which demonstrates the superiority of the cGAN-based approach in gen erating more human recognisable images. In conclusion, the thesis demonstrates the effectiveness of conditional GAN-based meth ods for image superresolution. The results validate the effectiveness of method in generat ing high-quality, visually consistent high-resolution images from lower-resolution inputs. The findings highlight the importance of incorporating class labels in the form of em bedding layer in the discriminator networks. Future research can explore different loss functions, network architectures, and larger datasets for further validation. The proposed method have potential for practical applications in various fields.