Most existing dehazing algorithms suffer from under-or over-enhancement, color distortion, and halo artifacts. An improved method of atmospheric light estimation using quad-tree subdivision and an improved guided filter were proposed to solve these problems. First, a more faithful estimate of global atmospheric light was produced by quad-tree subdivision using a non-overlapped dark channel. Then, the reasons for the existence of halo artifacts in edge regions were discussed and an adaptive weight was added to the guided image filter. The improved guided image filter was used to refine the raw transmission map. Finally, based on the atmospheric scattering model, a dehazed image was obtained using the estimated atmospheric light value and refined transmission map. Experimental results indicate that the color of the dehazed image is more reliable and halo artifacts in edge regions are reduced. The proposed algorithm performs better than state-of-the-art haze removal algorithms in terms of color fidelity and detail enhancement.