In real life, fog is generated under specific weather conditions, which reduces the color fidelity and information integrity of outdoor images. Although dehazing methods based on the convolutional neural network (CNN) had been rapid and significant progress, there are still problems of non-homogeneous haze that cannot be completely removed in the image, and the image after dehazing appears conspicuous color and structure migration between the hazy-free image and the clean image, which cannot be consistent with the human visual system. Based on the above problems, this paper proposes a dehazing model with encoder-decoder architecture, which enhances the feature extraction ability of non-homogeneous by embedding channel attention and pixel attention in the residual blocks, enhancing the dehazing performance of the model. At the same time, the multiscale structural similarity index (MS-SSIM) loss function and Mean Absolute Error (MAE) loss function is introduced to make inconspicuous color and structure deviation between the dehazing image and the clean image. Experiment results show that the Peak Signal to Noise Ratio (PSNR) is improved by 1.27% without reducing the Structural Similarity (SSIM). The image after dehazing is more matched with the human visual system, which effectively solves the problem of incomplete image dehazing.