A Conditional Generative Adversarial Network Model for Sketch-to-Image Translation

Abstract

The sketch-to-image translation is a challenging task that involves generating realistic photographs from hand-drawn sketches. While Generative Adversarial Networks (GANs) have achieved remarkable success in generating real images, sketch-to-image translation remains difficult due to the limited information in the input sketches. Generative modeling problems can be addressed using GANs, which are a category of artificial intelligence algorithms. The objective of a generative model is to discover the probability distribution by analyzing training samples. This estimated probability distribution is then utilized by GANs to produce additional instances. In this paper, we propose a conditional GAN-based model for sketch-to-image translation tasks. By incorporating a loss function in the training process and learning mappings from input sketches to output images, these networks can handle various problems that previously required separate loss formulations. This approach allows for a versatile problem-solving technique using the same fundamental principles. Our suggested approach involves adversarial training of a generator and a discriminator, both of which are integrated into our model. The generator network takes the sketch and a textual description as input and generates a corresponding image. The discriminator network within our model is trained to distinguish between generated and authentic images. Consequently, the generator network is trained to produce images that can deceive the discriminator network into perceiving them as genuine.