Abstract: Based on the instance segmentation model of Mask R-CNN, this paper proposes a new deep learning method named MixedMask. It offered two effective strategies. (1) The convolution kernel with mixed scale was used to improve the network's capability of extracting low-resolution instances. (2) Based on the squeeze-and-excitation networks, the improvement was made to solve the problem of channel information loss caused by the dimension reduction in the original network. Test on balloon datasets and xBD datasets shows that this method reaches 83.46% and 58.92% AP (IoU=50) respectively. Compared with the Mask R-CNN, the results were increase by 1.3% and 5.9% respectively.