This work demonstrates the potential of the proposed CycleGAN (cycle generative adversarial network)-inspired model in cross-domain image translation tasks for ADAS (advanced driving assistance systems). The proposed model consists of two generator networks for translating images between cross-domains and two discriminator networks that distinguish real images from the translated images. The model has been trained to minimize the formulated loss function, i.e., reconstruction loss, which encourages the tailored generator networks to learn shared representation between the two domains to effectively facilitate cross-domain translation. The proposed model performs the cross-domain image translation with feature preservation. The effectiveness of the proposed model and the learning algorithm has been validated through three cross-domain image translation datasets. The FID score improved by up to \(50\%\) , and the LPIPS score improved by up to \(30\%\) in the various tests. Consequently, experimental results demonstrated that the model outperforms the baseline CycleGAN model, translating images from one domain to another on three datasets while preserving features.