Revamped Lane detection in hazy environments using Denoising Diffusion Probabilistic Model
Sivakumar Palanirajan, B Natarajan, Arun Baalaji, Ebin Joe · 2024
In today’s world of rapidly growing technological markets, image processing has become vital for systems to work on visual data. A noteworthy task would be lane detection in autonomous vehicles. A good lane detection system should be able to perform the processing irrespective of the environment and this mandates that adequate preprocessing should be done. One such preprocessing step is image dehazing. Dehazing is particularly tricky with a heavily hazed image where the background object’s visibility is less than zero. To address this issue, methods like the Atmospheric Scattering Model (ASM) have been constructed to dehaze the image. However, these methods tend to lose the color details of the image. The Denoising Diffusion Probabilistic Model (DDPM) has recently emerged as a tool for image generation, showing promise in addressing this issue. This research work aims to build a hybrid model by combining the DDPM and ASM approaches named RevampedDehazingDDPM (RD-DDPM). The proposed model recreates the original image while maintaining the visual details to the plausible limit. The proposed model produced PNSR score of 38.59 dB, SSIM score of 0.0989 and LPIPS score of 0.0040 which showcases the outstanding performance of the proposed model.