The presence of smoke can cause the damage or loss of image target information. In view of the local nature of smoke in the scene, a smoke removal precheck mechanism based on the target detection Yolov3 algorithm is proposed in this work, that is, a precheck mechanism is added in the smoke removal process to realize the directional removal of smoke on the smoke image, improve the efficiency of smoke removal and avoid the impact of smoke on the non-smoking area. Different from the existing deep learning-based defogging methods for visible images, this method takes four polarization images as network input, and uses multi-scale attention adversarial network to extract the polarization information of the target in the smoke area, so as to alleviate distortion and enrich the structure and detail information of the target after smoke removal. Qualitative and quantitative experimental results on real data sets show that the proposed algorithm can effectively improve the smoke removal effectiveness and efficiency of polarized images.