One of the toughest challenges for AI in segmentation involves preserving narrow lines, delicate boundaries, and elongated features that cover only a small part of the image. While deep neural networks can understand broader context through multiple layers, they often lose fine spatial details. Losing just a few pixels along a linear feature can mean losing the entire structure. This patent enhances preservation of both large and small details by combining global context with local features and adding an edge-detection network. Edge data is gradually integrated into the decoder during reconstruction, giving the segmentation process insights into feature presence and exact boundary locations. This method is especially effective for images requiring precise segmentation of linear structures, not just detection, addressing a key challenge: high-level networks may miss small but vital spatial cues for thin objects. Focusing only on edges might ignore context, which is important for understanding the significance of different boundaries. The patented design fuses multiple scales and evidence types during decoding, showing how representations can be adapted to various datasets and geometric details. For certain tasks, the most critical information for AI is exactly what traditional methods often overlook.
Helping AI maintain fine structures in images
Patent number: CN 118865171 A
Inventor(s): L. Bai, T. Zhang, J. Liang, C. Xia, N. Wu, and D. Mo
Citation: L. Bai, T. Zhang, J. Liang, C. Xia, N. Wu, and D. Mo, “图像中线状结构识别分割的深度学习模型、方法、存储介质和装置 [Deep-learning model, method, storage medium and apparatus for recognition and segmentation of linear structures in images],” China Patent Application CN 118865171 A, Oct. 29, 2024.
