The sky region of restored images often appears serious noise and color distortion using classical dark channel prior algorithm. To address this is- sue, we propose an improved dark channel prior algorithm which recognizes the sky regions in hazy image by gradient threshold combined with the absolute value of the difference of atmospheric light and dark channel. And then we es- timate the transmission in sky and non-sky regions separately. At last, we en- hance the brightness and contrast of results. Experimental results show that our restored images are more natural and smooth in sky regions.