IMAGE ENHANCEMENT OF WEATHER DEGRADED IMAGES USING COMPUTER VISION TECHNIQUES

Abstract

Environmental effects, mist, haze, fog, snow and rain considerably effect visibility. The poor quality weather-degraded images perpetually effects performance of automated surveillance and tracking systems. Images enhancement and retrieval has a wide range of application, such as tracking and surveillance systems, consumer electronics and autonomous robotics Systems. Such applications generally require computationally efficient algorithm for cost effectiveness. By studying visual manifestation of various weather conditions in images, the environmental characteristics can be modelled effectively. Water droplets present in atmosphere cause mist, fog and haze effects due to scattering of light as it propagates through these particles. Subsequently, chromatic effects of image scattering can be reversed for retrieval of image information. Scattering of light affects image contents in proportion to the depth of scene. Classical image enhancement procedures are not effective as these do not take into account depth of image. Various model based as well as computer vision based techniques to enhance quality of weather-degraded images are in vogue. Physical modelling approaches, although promising better results in terms of color fidelity and contrast are computationally very expensive. Computer vision based techniques are computationally efficient; however usually result in compromising important information. An optimum blend of these two techniques is generally considered efficient means for the desired solution. Single image dehazing technique using dark channel prior is an advance approach with computational advantage being of first order. This technique has been further refined in this thesis by further improving estimation of Atmospheric Light and optimizing transmission sensitivity of the model. Contrast of the restored images has been considerably improved vis-à-vis color fidelity further refined. Improved re-defined model provides even better control on restored image parameters and fine tuning of contrast and color fidelity of recovered images. Feature detection, cross correlation, image registration, matching and recognition improves as input image quality improves. Image dehazing is an added feature to latest Night Vision Devices which can even pay dividend if utilised as a pre-processing stage. Major application of Automated single image dehazing techniques also include pre-processing of UAVs, GIS, and satellite imagery, where it is not feasible to obtain same images again.