Multi-resolution parallel ResNet（MPR） and Attention FPN（AFPN）are presented based Mask R-CNN（MA Mask R-CNN） to achieve image instance segmentation,which aims at improving the feature extraction ability of the model,and efficiently increasing the detection and segmentation accuracy.MPR adds parallel branches to extract the feature before each down-sampling of ResNet to make full use of the features.Attention FPN（AFPN） learns the attention mechanism by adding attention modules to each layer of the FPN,so that each layer can adaptively adjust the weight,emphasize the effective information,and suppress the unimportant information.Results of the comparison of the proposed MA Mask R-CNN to the previous state-of-the-art models on COCO dataset show a wider variety of scenarios and complete labels.Our method in instance segmentation task performs better,compared with the state-of-the-art methods.m AP,m AP50 and m AP75 are improved by 1.1%,1.6% and 1.3 respectively,to Mask R-CNN,and 0.4%,1.3%,0.4% respectively,to PANet,and 7.4%,7.9%,9.5% respectively,to FCIS.The average segmentation accuracy of small,medium,and large targets（m APS,m APM,m APL） is increased by 1.4%,1.0%,and 1.1% to Mask R-CNN,respectively.