There exist today plenty of algorithms and many papers about dehazing or defogging, that is enhancing images taken in hazy or foggy conditions. To our knowledge none of them has got a signifcant result for dense and non-dense haze image at the same time. In this master thesis, we will propose an algorithm that are able to dehaze both dense and non-dense hazy images. Our hope comes from the fact that we observe in dense hazy images, the Spatio-Temporal Retinex-inspired with Stochastic Sampling (STRESS) framework (Kolas et al., JIST, 55(4), 2011) gives a more visually pleasing result when we compare it with the DCP algorithm (He et al., CVPR, 2009) for the same input. Our hypothesis is that STRESS uses one or more of its principles to enhance effciently dense hazy images. In this work, we will show how we find out this hypothesis and also justify it. For the purpose of our experiment, we define a new database where images are separeted according to their degree of haziness (fuzzy, medium and very fuzzy). The dehazing algorithms that we consider are typically (Fattal, Proc. ACM SIGGRAPH, 27(3), 2008), (Fattal, Proc. ACM SIGGRAPH, 34(1), 2014), To evaluate the quality of these dehazed images, we use some metrics and a psychophysical approach. From this experiment on the previous works, we show their relationships with STRESS and finally the performance of the new algorithm which is the combination with some of them with STRESS idea is showed as well. Basically our algorithm assesses the hidden free-haze layer by assuming that there are three patterns almost in outdoor hazy images namely: sky region (or regions which have the same behaviour, like snow), far objects and near objects. The experiment shows also that our algorithm based on the two-scale STRESS approach, edge detection and Hidden Markov Model has often more visibility level than most of state-of-the-art algorithm when using the metrics de- fined in (Hautiere et al., Image Analysis & Stereology Journal, 27(2), 2008).