Hazy conditions significantly reduce image contrast and obscure object boundaries, impairing the performance of vision-based tasks such as object detection, tracking, and scene understanding. Learning-based dehazing methods have attained numerous achievements in dehazing images. For real-world haze images, the current methods result in the poor quality of haze-free images. In this study, we propose an image dehazing method based on cycle generative adversarial network (CycleGAN), which integrates the transmission map and depth estimation (CycleGAN-TMDE). In CycleGAN-TMDE, we designed a dehaze generator that includes a transmission map estimator and an atmospheric scattering model to produce haze-free images with real-world physical characteristics. To further improve the dehaze generator’s dehazing capability, we adopt a depth estimator to generate haze images while simultaneously using the dehaze generator to remove haze from these generated images. The cycle loss function compensates for the absence of matched hazy sample pairs in unsupervised learning. The adaptive loss function enhances the model’s robustness, ensuring that when a haze-free image is used as input, CycleGAN-TMDE can produce similarly clear outputs. On the real-world hazy images of the realistic single image dehazing (RESIDE) dataset, CycleGAN-TMDE achieves clearer and more natural haze-free images, particularly producing better visual effects for distant scenery while also yielding favorable no-reference image quality assessment (NIQE) metrics. On the synthetic hazy datasets RESIDE and Haze4k, CycleGAN-TMDE can restore high-quality haze-free images while achieving comparable peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) values to supervised learning methods and outperforms other unsupervised learning methods. The code for the method in this article can be downloaded at
https://github.com/Paris0703/CycleGAN-TMDE.
DownLoad: