Instance segmentation is of great significance for eliminating safety hazards brought by irregular machinery and equipment on construction sites and for monitoring workers. However, the current mainstream instance segmentation models have the problem of low boundary detection accuracy. Combining the characteristics of instance segmentation, this paper proposes an improved Mask R-CNN model of multi-stage refining mask based on the global context channel attention (GCCA) mechanism. First, this paper gradually fuses finegrained features in the mask head in a multi-stage manner to refine high quality masks. Second, in order to better fuse fine-grained features, a GCCA attention mechanism is constructed, which aggregates global features through a simplified global context module, and utilizes onedimensional convolution to achieve local channel interactions without dimensionality reduction. The experimental results show that this paper has achieved great results on both COCO and MOCS datasets. Among them, compared with the Mask R-CNN model, the average accuracy of the algorithm in this paper in detection and segmentation is improved by 2.4% and 7.6% respectively.