Piecewise linear model for haze level estimation and an efficient image restoration technique

In this paper, a piecewise linear predication model that relates light extinction coefficient (bext) to log of image contrast/transmission is proposed. The objective is to calculate more accurate bext values compared with previous treatments which used linear model of contrast relationship only. A multivariate linear regression model is learned for each linear portion of the model. Results show superior behavior of the proposed model in terms of error calculated between estimated values and ground truth values. Moreover, an effective approach to dehaze and enhance outdoor images captured under adverse weather conditions is introduced. A raw transmission map is estimated using a dark channel prior, then guided image filtering is used in two stages; first for transmission map refinement, second for details enhancement. Guided filtering not only exhibits the edge-preserving smoothing, but also a structure transferring property. Details enhancement results in improved visibility with minimum information loss.