Traditional methods of image dehazing can distort color in areas such as the sky, white clouds, and bright areas. To address these problems, a three-step method is proposed for removing image hazing using a multi-scale convolutional neural network (MCNN) and classification statistics. First, the MCNN is used to estimate the transmittance of the image. Second, the estimated transmittance is classified and the pixel values of the sky, white clouds, and other bright regions in the dark channel are determined. Finally, the radiance of the scene is smoothed by a low-pass Gaussian filter to produce a restored haze-free image. Experimental results show that this method preserves the color in bright areas after the image is defogged, retaining the natural appearance of the image. The proposed method achieves improved dehazing on both synthetic and real images.