Traditional retinal vessel segmentation methods often face challenges such as missegmentation caused by optic disc confusion, lack of continuity in segmentation results, and imprecise segmentation in detailed regions. To address these issues, a retinal vessel segmentation algorithm was proposed based on UNet. The algorithm replaced traditional square convolutions with a fusion of horizontal and vertical one-dimensional convolutions and two-dimensional square convolutions, enhancing the representation capability of the eye region. A multi-scale branch approach was adopted to increase feature space diversity, thereby improving the network’s feature learning and expression capabilities. Additionally, to further enhance segmentation performance, multi-layer dilated convolutions was introduced into the deep structure of the autoencoder, replacing traditional simple pooling operations. This approach enlarged the convolution kernel size and expanded the receptive field, achieving a fusion of multi-scale shallow and deep feature information. The proposed algorithm was evaluated on the public DRIVE and CHASE_DB1 datasets. Experimental results demonstrates that the algorithm achieves precision (0.956 8 and 0.959 8) and F₁ scores (0.832 6 and 0.830 4), respectively. Compared with traditional UNet and recent UNet-based retinal vessel segmentation methods, the proposed algorithm shows advantages in accuracy, sensitivity, specificity, and F₁ metrics, these validation results fully demonstrate the proposed model’s strong capability in precise segmentation tasks.