Named entity recognition (NER) is a fundamental task in information extraction and aims to locate the boundaries of entities in a sentence and classify them. In response to the fuzzy boundaries of nested entities based on span detection models, this study proposes a nested NER model based on span boundary perception. Firstly, it utilizes a bidirectional affine attention mechanism to capture the semantic relevance among word tokens and then generates a span semantic representation matrix. Secondly, it designs a second-order diagonal neighborhood difference operator and establishes a span semantic difference mechanism to extract semantic difference information among spans. Additionally, a span boundary perception mechanism is introduced to employ the local feature extraction ability of sliding windows to enhance the span boundary semantic differences, thereby accurately locating the entity span. The model is validated on three benchmark datasets of ACE04, ACE05, and Genia. The experimental results show that the proposed model outperforms related work in entity recognition accuracy. Additionally, the study conducts ablation experiments and case studies to verify the effectiveness of the proposed semantic difference mechanism and span boundary perception mechanism, providing new ideas and empirical evidence for further research on NER.