The clinical assessment of multiple cardiovascular illnesses necessitates the accurate segmentation of cardiac magnetic resonance (CMR) images. CMR images have the inherent defect of weakly-defined edge contours, and current methods tend to use samples from a single clinical center, which performs poorly when applied to MRI scanners at various clinical centers. This paper proposes a multi-center CMR image segmentation method based on enhanced DeeplabV3+ to address this issue. Based on the DeepLabV3+ segmentation network, the backbone network is replaced with MobileNetv2 and boundary extraction and feature fusion algorithms are introduced for the CMR images segmentation. The MICCAI 2020 M&Ms Challenge dataset is used to assess the approach, have the highest average Dice scores (0.8781), which improves the segmentation accuracy of the multicenter samples relative to the other segmentation models and makes them more valuable for practical applications.