Image dehazing based on dark channel has good performance and has been widely applied in computer vision. But with high computational complexity, it's difficult to apply this algorithm into real-time image dehazing. In order to realize the real-time HD image dehazing, a GPU-accelerated parallel computing implementation is proposed to optimize the image dehazing algorithm based on dark channel. A parallel computing implementation based on transposition is presented to optimize the mean filter in guided filter. Texture memory is used to accelerate the process of min filter. Reduction algorithm is used to calculate the value of atmospheric light of a haze image at high speed. Experiment results indicate that, under the premise of zero accuracy loss, the algorithm can deal with the 1080p with 38 frames per second and the speedup up to 51x to meet the requirements of real-time HD image dehazing.