Development and performance evaluation of enhanced image dehazing method using deep learning networks

In this work, a deep learning-based high-performance image dehazing technique is proposed for image processing applications. The end-to-end network model is constructed and implemented using a dehazing network, discriminator network, and fine-tuning network models. These three methods are well-trained individually using appropriate datasets. The individual network models are integrated as an end-to-end network model to enhance the dehazing process. The applied input hazy image is processed by the dehazing network model using the estimation of transmission map and atmospheric light along with parallel convolution layers. A discriminated dehazing image was extracted from the discrimination network. Finally, fine-tuning is carried out based on the results of the discriminator network model. Various hazy images from different datasets are collected and applied to the proposed model, and performance metrics such as PSNR, SSIM, and MSE are evaluated. Qualitative, and quantitative comparison and analysis are carried out between the proposed learning-based image dehazing and existing dehazing methods. The average PSNR value of the proposed dehazing model is obtained by a maximum of 40.7 %, and a minimum of 1.34 %, when compared to the existing works. The average SSIM of the proposed work is increased by a maximum of 22.12 %, and a minimum of 3.38 % with respect to the existing works. The maximum average value of MSE for the proposed model is decreased by 72.6 % and the minimum decrease of MSE is 4.08 % when compared to the state-of-art-works.