Contribuições à restauração automática de imagens de cenas em meio participativo

Abstract

This work deals with the problem of image restoration of monocular sequences of images acquired in participating media, i.e. media that interfere with light propagation. Specifically, the proposed work focuses on the automatic restoration of images acquired in underwater and hazy scenes, thus enabling the prompt use of legacy Computer Vision and Image Processing algorithms. The proposed restoration process requires a sequence of images, i.e. at least a pair of images from the same scene as input and produces as output an image sequence in which the participating medium effects are attenuated and the visibility improved. For underwater scenes, we propose an extension of the Dark Channel Prior (DCP) approach which uses a single image as input, which we call Underwater DCP (UDCP). In UDCP, the blue and the green color channels are used as the main underwater visual information source. The main limitation refers to images acquired from scenes containing objects with white or gray objects surfaces. Therefore, UDCP is only applied as an initial step of our method, which is based on an optical model of the participating media and three-dimensional reconstruction. This allows the automatic estimation of the medium parameters and the scene depth, which enables the inversion of the model, and thus to restore the images. Differently from previous works, our method does not need additional equipment or information -- only a intrinsically calibrated camera. The adopted optical model is an adaptation of the model proposed by Jaffe-McGlamery, which in turn was based on the one initially introduced by Koschmieder. This model considers the main degradation effects like absorption and scattering.We further use our method for three-dimensional reconstruction using structure-from-motion approach. The first step is obtaining the pixel correlation between consecutive images in the sequence. We then derive a novel formulation to estimate the optical flow problem in participating media called Generalized Optical Flow Model (GOFM). This approach generalizes assumptions of the Horn-Schunk model in order to tackle both non-participating and participating media. In the case of non-participating media, the model is simplified in the classical approach. The GOFM methodology associated with UDCP was evaluated both in a qualitatively and quantitatively way. It obtains better results than the combination of image restoration and classical optical flow. Experimental results in real and simulated environments were carried out in order to validate the proposed methodology. Taking into account the inherent difficulties of the participating medium, the results show that our method is able to adequately recover the attenuation parameter and restore the images. The restoration thus obtained provides better results than the state-of-art as evaluated by well defined.