Normalizing images in various weather and lighting conditions using ColorPix2Pix generative adversarial network.

Autonomous vehicles (AVs) are widely regarded as the future of transportation due to their tremendous benefits and user comfort. However, the AVs have been struggling with very crucial challenges, such as achieving reliable accuracy in object detection as well as faster computation required for quick decision-making. In recent years, perception systems in driverless cars have been significantly enhanced, mainly due to advances in deep-learning-based object detection systems. However, these perception systems are still heavily affected by environmental variables, such as changes in illumination, refractive interference, and adverse weather conditions, which may compromise their reliability and safety. This research proposes an advanced colour vision technique and introduces an efficient algorithm called ColorPix2Pix for normalizing images captured in various hazardous environmental and lighting conditions. Optimized Generative Adversarial Network (GAN) models were employed to address these challenges by normalizing images affected by extreme lighting and weather conditions. Specifically, a novel model, the ColorPix2Pix GAN, was proposed, incorporating an enhanced loss function designed to prioritize both structural and color fidelity. Through a custom loss calculation, the GAN is guided to reconstruct images more effectively, reducing noise and restoring essential visual details while maintaining natural color profiles. The methodology included a two-phase training process in which two comprehensive datasets were curated, capturing diverse environmental scenarios, such as fog, rain, and variable illumination conditions, to simulate real-world challenges for AVs. These datasets enabled robust training and testing of the ColorPix2Pix model to assess consistent performance across complex scenarios. Furthermore, a custom loss function combining perceptual loss with color consistency measures was used to increase the GAN's sensitivity to both structural accuracy and color fidelity, allowing the model to simulate the natural appearance of original images while effectively removing environmental noise. The model was trained under varying adverse conditions to enhance adaptability. For the lighting dataset, it achieved an average SSIM of 0.767, PSNR of 68.581, LOE of 0.171, and LPIPS of 0.205. On the weather dataset, it recorded an SSIM of 0.660, PSNR of 67.185, LOE of 0.177, and LPIPS of 0.241. These results highlight the model's effectiveness in restoring image quality and improving perceptual reliability for autonomous vehicle systems, outperforming existing image normalization methods.