A Comparative Analysis of Image Dehazing using Image Processing and Deep Learning Techniques

The haziness presence in atmosphere now poses to be a critical issue, especially in terms of visibility and visual quality. The major challenge faced in the present day scenario is to stitch a technique that detects the presence of haziness and to provide an efficient method to remove it as well. This process is known as image dehazing. Although different dehazing algorithms has been introduced, still better performances can be achieved which is addressed in this paper. This challenge is overcome by using a suitable classification algorithm to detect hazy images and then cleaning it to get corresponding clear image. In this method, two goals are presented and executed, Initial one be a comparative analysis between the machine learning and deep learning models is presented for the classification of hazy and clear images and then the hazy images are processed by a robust deep neural network for the removal process. In order to classify, two ML algorithms, namely, SVM and Logistic Regression are used and two CNN standard models Resnet, Vgg are also used and their performances are compared to obtain best classification algorithm for this problem. In classification model, standard CNN models outperforms Machine Learning models by a fair margin by proving to be more accurate. Similarly for dehazing also DCP Image processing algorithm is implemented with a proposed CNN de-hazing network and their performances are also compared to prove the efficiency of proposed CNN model. ITS, HSTS, Indoor SOTS Dataset are used as inputs. The performance of CNN network is compared with an image processing technique called Dark Channel Prior (DCP) algorithm. Comparatively the CNN model outperforms the DCP algorithm by attaining higher image quality metrics. To evaluate the image quality, two metrics are compared, they are Structural-similarity-index (SSIM) and peak-signal-to-noise-ratio (PSNR).