"U-Net: Convolutional Networks for Biomedical Image Segmentation" by Ronneberger et al. introduced the U-Net architecture that's now standard in medical image analysis. The paper describes an elegant encoder-decoder network structure that preserves spatial information critical for precise segmentation. The architecture has influenced numerous subsequent designs and remains relevant for current segmentation tasks.
"Mask R-CNN" by He et al. extended the Faster R-CNN object detection framework to include precise instance segmentation. The paper presents a straightforward yet effective approach to segmenting individual objects while maintaining real-time performance. Its implementation has become a cornerstone for modern instance segmentation systems.
"DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Pooling, and Fully Connected CRFs" by Chen et al. introduced atrous convolution for dense feature extraction. The paper shows how to maintain high resolution feature maps without excessive computational cost. Their approach significantly improved segmentation accuracy while keeping reasonable processing times.