The task of Named Entity Recognition (NER) is an important component of information extraction tasks. Currently, span-based approaches are receiving widespread research attention. Despite their success in many aspects, these approaches also suffer from insufficient utilization of non-entity information. In this work, we view the span-based NER task as an edge detection task and propose EDNER (Edge Detection for Named Entity Recognition). In this method, we define the edge representation of text and generate deep representations through a specially designed edge interaction layer, while auxiliary edge contrastive learning and global edge detection task allow the model to fully adapt to the task of edge detection. Our carefully designed convolutional layers and prediction layers extract entity spans by detecting entity edges. Experimental results show that our method outperforms existing state-of-the-art methods and achieves highest F1 scores on four benchmark NER datasets.