An Improved Small Feature Segmentation Algorithm Based on U-Net Model

The segmentation network of small features can be used in the medical field, such as the segmentation of retinal blood vessels, which can help doctors diagnose eye related diseases. However, the current retinal vessel segmentation methods still lack the ability to segment microvessels. In this article, we propose an improved small feature segmentation algorithm for retinal vessel segmentation. Based on the U-Net algorithm, improvements are made to address the characteristics of free flow and rich morphology of retinal blood vessels. Based on a multi-scale dilated convolution module, multi-scale information of retinal blood vessels under different receptive fields is aggregated to enhance microvascular feature extraction. Experiments were conducted on a mixed public dataset. The results showed that our method can accurately identify and segment retinal microvessels, with better segmentation performance. What’s more, the proposed model can also be applied to small feature segmentation in industrial fields, such as crack detection of large-size ultra-thin photovoltaic silicon wafers.