Abstract: Transparent objects are common things in daily life and have unique visual characteristics. Sensors to accurately estimate their depth. In most cases, the depth information captured by visual 3D sensors appears transparent, the depth value of the background behind the object or a large area of depth loss. To quickly repair the depth loss of transparent objects in the depth image, a method for rapid repair of the depth image of transparent objects based on semantic segmentation and multi-scale fusion was proposed. Firstly, predicted the mask of the transparent object by light real-time semantic segmentation. Secondly, we removed the wrong depth information in the mask fields of the depth scene image. Finally, we performed multi-scale feature extraction and feature fusion on the color image and the error-removed depth image. Quickly completed repair the depth image of transparent objects. Experimental results show that the proposed method achieves results of 0.027, 72.98, and 98.04 on the measurement MAE, δ_1.05, and δ_1.25 for the depth repairment of transparent objects, which are better than the existing methods.