In micro-action recognition, accurately locating coarse-grained action boundaries and capturing detailed fine-grained actions are both crucial. We introduce DualActNet, which integrates the benefits of the SlowFast network architecture with advanced attention mechanisms. The slow channel employs HaloNet Attention to enhance local information processing and delineate coarse-grained actions. Conversely, the fast channel utilizes CBAM (Convolutional Block Attention Module) Attention, which merges spatial and channel attention to heighten sensitivity to fine-grained actions. In the fusion stage, SE (Squeeze-and-Excitation) Attention re-weights channel features to integrate insights from both channels synergistically. A dynamic loss function is introduced to address the class imbalance. Testing on the MA-52 dataset, DualActNet demonstrates superior performance over the baseline SlowFast and the MANet models, achieving higher \(F1_{mean}\) and Acc-Top1 metrics for both action types.