EFFICIENT IMAGE SYNTHESIS USING GENETIC ALGORITHM – GENERATIVE ADVERSARIAL NETWORK

Abstract

GAN’s have emerged as a powerful technique for generating high-quality images due to their unique characteristics and capabilities. In this report, there is a discussion for the motivation behind using GANs over other generative models for example Restricted Boltzmann Machines (RBMs), Deep Belief Networks (DBMs), and Variational Autoencoders (VAEs). This research highlights the advantages of GAN’s in terms of image quality generation. To gain a comprehensive understanding of GAN’s and their practical implementations, several studies have been conducted that aided in the creation of a GA-GAN framework. This research provides insights into the theoretical foundations and practical considerations of GANs for image synthesis. This paper introduces unsupervised learning techniques specifically designed for GANs, enabling their effective utilization with small datasets such as MNIST and CIFAR-10. Driven by the knowledge gained from these resources, this report shows a novel implementation on Genetic algorithm-based GAN model which are supported by learning rate schedulers. The approach incorporates various essential concepts and techniques to enhance the quality of image generation using limited datasets. Specifically, methods like normalization, data augmentation, batch normalization, and Adam optimizer were used to enhance the overall accuracy of GAN model. However, this report uses genetic algorithm instead of gradient based approach and also generate high quality image than the real images. For these experiments, the Anime Face Dataset was collected from Kaggle through API integration. This dataset comprises approximately 63,565 anime face images, which is similar in scale to the widely used CIFAR-10 dataset. By employing GA-GAN model with genetic algorithm for optimization, this research work aims to generate high-quality anime face images. The proposed framework employs two performance metrics termed as real_score of 0.9722 and fake_score of 0.0452. Binary cross entropy loss function was used for both generator and discriminator. These metrics provide valuable insights into the quality and diversity of the generated images. Additionally, the Fréchet inception distance (FID) score was also discussed, which is a widely used metric for evaluating the quality of generated images. The FID score compares the feature embeddings of the generated images and the original dataset using a pre-trained Inception model. A lower FID score indicates a closer similarity among original images and generated images, highlighting the success of the GAN network in capturing the underlying data distribution