INTEGRATED CLASSIFICATION AND ENHANCEMENT OF LOW LIGHT HAZY AND NOISY IMAGES

Abstract

Images captured in low light environment losses out on details due to poor visibility conditions. These images are further degraded by noise and haze which result in contrast reduction, low visibility and color fading. The problem of haze removal is complicated since there is an inherent ambiguity between underlying scene and haze. Similarly, noise contained in an image may get ampli ed during haze removal. Hence, these unnecessary artifacts (haze and noise) are required to be removed before image enhancement in order to make non uniformly illuminated noisy and hazy images suitable for computer vision and multimedia based applications which are primarily designed for high-quality inputs. This thesis proposes a novel integrated algorithm wherein image classi cation into Hazy, noisy and non-noisy is carried out automatically along with image enhancement. Since, the noise level in an image is not precisely known a priori, classi cation of im- age as noisy or non-noisy may lead to erroneous outcomes. Hence, image classi cation based on haze is carried out rst. Hazy images are classi ed on the basis of visibility Parameter. Images with visibility parameter value lower than a threshold of 1.16 are treated as hazy and higher than 1.16 are treated as non-hazy. Non Hazy images are then tested for noise. Support Vector Machine classi er is used for classi cation of images into noisy and non-noisy category. After classi cation of images as hazy, noisy and non-noisy, di erent approaches are adopted for image enhancement. Multilayer Perceptron of Arti- cial Neural Network classes is used for haze removal. Noisy images are enhanced using Retinex based method which utilize image decomposition into re ectance and illumina- tion component through iterative approach. Multi-Layer lightness statistics is used for enhancement of images which are neither hazy nor noisy. A wide variety of low light noisy and hazy images are selected to evaluate the perfor- mance of proposed algorithm. Qualitative human evaluation with various state of the art and modern techniques of image enhancement is carried out, so that an assessment can be built for comparisons. Mean Opinion score is used for Qualitative evaluation. The results are further strengthen by Quantitative evaluation using various methods such as Lightness order error, Structure Similarity, Naturalness Image Quality Evaluator and No reference Free Energy based Robust Metric. The method of image classi cation into hazy and noisy using visibility Parameter and Support Vector Machine Classi er is novel and works well for most of the images. The comparisons results presented in chapter 4 indicates that integrated approach has performed similar or better than other algorithms for enhancement of noisy and hazy images.