Image Super-Resolution (SR) focuses on converting low resolution images into high resolution images. Recently, numerous studies apply Transformer to SR tasks, demonstrating significant effectiveness. However, these methods are constrained by a limited receptive field, which hinders the utilization of low resolution image feature information. Thus, it impedes the comprehensive exploitation of the Transformer’s multi-level contextual extraction capabilities. To overcome this limitation, we introduce a Dual-Level Attention network (DLAFormer). DLAFormer integrates channel spatial attention with window based multi head self attention, effectively exploring the relationships of pixels in terms of spatial, channel, and local features. This adaptive mechanism enhances the extraction of critical information, resulting in higher resolution reconstruction outcomes. Additionally, we incorporate Deformable Convolution (DC) to adjust and optimize the attention weights, improving the adaptability and sensitivity of model to local features such as edges and textures. We conduct extensive comparative and ablation experiments, showing that our network outperforms benchmark methods.