To address semantic drift, boundary blurring, and detail degradation that commonly arise in resource-constrained video portrait matting, especially with complex boundaries and long temporal sequences, this study proposes an efficient method based on dual-branch encoding and semantic guidance. In the encoding stage, a dual-branch structure with high-bit and low-bit branches is constructed. The low-bit branch performs fast feature extraction for the majority of video frames, thus reducing computational complexity and storage overhead. Meanwhile, a high-bit keyframe branch is periodically activated to refresh the semantic representation, alleviating the limited expressive capacity of long-term binarized modeling. In the decoding stage, a lightweight module for keyframe semantic guidance and region-adaptive gated fusion is designed. This module explicitly models the complementary relationship between keyframe semantics and information propagated from the previous frame in the feature space. As a result, the model prioritizes temporal continuity in stable subject regions while enhancing detail sensitivity and structural consistency in boundary regions and regions with significant changes. Experimental results on multiple public video matting benchmark datasets demonstrate that the proposed method significantly reduces average computational cost and memory usage while effectively improving the temporal consistency and boundary quality of the matting results. In terms of structural design, the proposed method has the potential for further deployment optimization in resource-constrained environments such as mobile and edge devices.