A retinal vessel segmentation algorithm based on a lightweight high-frequency Transformer and feature complementary fusion was developed to address the issues in retinal vessel segmentation, including the loss of vessel information, the low proportion of vessel pixels, and the difficulties in segmenting vessels in lesion regions. A bilateral feature extraction encoding structure was constructed using the lightweight high-frequency Transformer module. Global and local information were effectively aggregated to mitigate the loss of vessel information during the initial encoding process. A detail-enhancement attention module was designed within the encoding stage to accurately distinguish the vessel foreground from the non-vessel background, thereby improving the model’s ability to capture fine vessels. A dual-decoding path was constructed at the decoding stage, and a feature complementary fusion module was introduced to fully integrate low-level details with high-level semantic information, thereby enhancing the segmentation accuracy of vessels in lesion regions. Experimental results on the public datasets of DRIVE, STARE, and CHASE_DB1 indicated that the proposed algorithm achieved accuracies of 97.12%, 97.62%, and 97.65%, sensitivities of 80.23%, 80.48%, and 81.30%, and F1-scores of 82.99%, 83.73%, and 81.73%, respectively. The parameter count and computational complexity (FLOPs) were 4.27 M and 1.79 G, respectively. While achieving high computational efficiency, this method maintains the model’s segmentation performance, fully demonstrating its potential in clinical diagnosis of ophthalmic diseases.