To address the problem that limited CSI dimensionality in embedded WiFi sensing scenarios leads to insufficient recognition accuracy for fine-grained alphabet gestures, a lightweight CSI-based fine-grained gesture recognition method is proposed. A staged preprocessing pipeline, including Hampel outlier removal, median filtering, wavelet denoising, and cubic spline interpolation, is applied to CSI sequences to improve signal quality and sequence consistency. Subsequently, a dual-branch network, termed DFANet, is designed to fuse time-frequency features by extracting subcarrier correlation features in the frequency domain and temporal dependency features in the time domain, and a dendritic network is employed for efficient classification. Experimental results show that the proposed model achieves average recognition accuracies of 88.51%, 92.64%, and 91.60% in laboratory, office, and classroom environments, respectively, with an overall average accuracy of 91.21%. The model contains 147.880×103 parameters and requires 49.534×106 FLOPs. Compared with ResNet50, under comparable recognition accuracy, the number of parameters is reduced from 21.291×106 to 147.880×103, and the computational complexity is reduced from 835.781×106 to 49.534×106 FLOPs. The results demonstrate that the proposed method achieves effective fine-grained alphabet gesture recognition while maintaining low computational cost and lightweight characteristics, providing methodological support for future deployment in resourceconstrained embedded WiFi sensing systems.