In recent years, deep learning has demonstrated excellent performance in the field of change detection. However, there are still shortcomings such as unclear fine-grained boundaries and internal voids caused by the loss of detailed information and insufficient semantic information in terms of false detections and missed detections. To address this issue, this paper proposes a new multi-scale adaptive fusion network（MSAF-Net）, which uses ResNet as the backbone network. Firstly, it presents a feature difference fusion module (FDFM), which can roughly extract boundary information from changed regions through simple operations. Then, to tackle the problem of lost detailed texture information, a multi-scale dual-feature fusion module (MDFM) is proposed to aggregate low-level texture information and high-level semantic information from multiple receptive fields. Furthermore, to solve the problem of insufficient high-level semantic information in complex scenes, a feature synergy enhancement module (FSEM) is put forward to extract deeper high-level semantic information by enhancing the fusion of feature information between branches. The integration of the FDFM, MDFM, and FSEM modules enables efficient and accurate pixel-level change recognition. The network adopts a dual-stream-single-branch architecture and improves performance through the following key technologies: performing difference processing in two stages, combining the shallow FDFM module to capture detailed differences and the deep MDFM module to conduct cross-scale feature fusion; and introducing the convolutional block attention module（CBAM）, in the feature fusion stage to adaptively weight important changed regions. The experimental results show that MSAF-Net achieved F1 values of 90.71%, 91.71%, and 82.71% on the LEVIR-CD, WHU-CD, and SYSU-CD datasets, respectively. Compared with the current mainstream methods, the performance has improved by 0.14, 1.2, and 0.87 percentage points, respectively.