An Efficient Image Dehazing Technique Using DSRGAN and VGG19

Haze and clouds can cause the weather to affect the clarity and contrast of photos that are taken by cameras, air quality, and other reasons. To address the issue of haze, image dehazing has become an area of significant importance. However, many current techniques for unsupervised picture dehazing rely on simplified atmospheric scattering models and a priori knowledge, which can result in inaccuracies and poor dehazing performance. The study of image-dehazing techniques has shown promise Generative Adversarial Networks (GANs) were developed. Unfortunately, because haze is so complicated, the bi-directional mappings domain translation techniques currently used in unsupervised GANs are not suitable for dehazing image work. With haze-free photos, the image may become distorted, lose image details, or retain image features poorly as a result. The paper makes recommendations for an end-to-end unsupervised image-dehazing system that is used to address these problems. Deep Super-Resolution Generative Adversarial Network (DSRGAN) that uses VGG19 for feature extraction. The problem addressed in this project is to create a novel picture dehazing method that uses the DSRGAN and VGG19 models to successfully remove haze from photographs. The specific objectives of this project are to implement the DSRGAN and VGG19 models, modify them for image dehazing, acquire a dataset of hazy images, preprocess the dataset to remove any anomalies, train the models on the dataset to learn the features and characteristics of hazy images and assess the performance of the models using both qualitative and quantitative metrics. The proposed method showed an improvement in both PSNR and SSIM metrics with 24.06 and 0.912. The proposed system achieves negligible generator loss and discriminator loss, offering a promising solution to the challenge of image dehazing in complex and dynamic environments.