The Car Image Generation Quality of DCGAN and VSGAN: A Comparative Study

In modern society, car image synthesis is used for many tasks. Vehicle identification and car image analysis are some uses. There are different types of Generative Adversarial Networks (GANs) for these tasks. For security and identification purposes car image generation is of utmost importance. DCGAN, being the most used, is a very popular choice. VSGAN is new, but is built for this specific purpose and CGAN is also utilized in car image classification. Prior works of various magnitudes are accounted. To see which model performs best is a very eccentric query. To find out, computing times, losses, accuracy and pictures generated of VSGAN, DCGAN and CGAN models with car image dataset are compared. The methodology is explained in detail. The architectures of the models are discussed along with their special and loss functions. The dataset images are already pre-processed and 6000 images are taken as input data. All models run 40 epochs for both training and testing phases with a set of fixed hyperparameters. For accuracy comparisons k-nearest neighbors (KNN), regression and random forest classifier (RFC) are used. This research aims to display the higher capabilities of the VSGAN in comparison to DCGAN and CGAN for this task. It can help in making better vehicle identification software and better image generation for forensics.