The current semantic segmentation models can already achieve good results in most scenarios. However, for dense object segmentation, most models have a significant drop in segmentation accuracy between dense object classes, especially for edge prediction, where adhesion or discontinuity will occur. To solve such problems, we present a neural network model with a dual-branch structure, named EBS-Seg. Apart from the main branch, additional net-work branches are introduced to focus on learning the edge of a single object and the overall edge topology information, then the final edge features and main branch features are fused. Moreover, we devise a loss function aimed at ensuring the efficient convergence of the edge learning component. To validate the effectiveness of EBS-Seg for dense segmentation tasks, we have incorporated a dense particle segmentation dataset derived from an industrial setting. Experimental results demonstrate that the segmentation performance of EBS-Seg exhibits superior edge continuity; at the same time, in terms of mIoU and F-score for edge, EBS-Seg can achieve the best performance over mainstream methods.