The formation pattern of hazy image based on multiple scattering, which is described by atmosphere point spread function (APSF), is modeled. The generalized Gaussian distribution was adopted to approximately formula the expression for APSF through the similarity in the shape and proper convolution operation. Aiming at enhancing the quality with reasonable time consuming, an improved dehazing method was proposed based on dark channel prior (DCP). The approach applies the superpixel algorithm to obtain image patches with similar depth for halo prevention, and based on this, the sky region was also detected more accurately by region merging. Therefore the transmission on the sky and non-sky region was separately estimated, the color distortion reduction in the sky region was achieved. Experimental results show that the proposed method can robustly recover a high-quality haze-free image with abundant details from both subjective and objective image-quality assessment.