parallel dilated convolutions at multiple rates that capture multi-scale context. This addresses a fundamental challenge in segmentation: objects appear at many different scales, and the network must recognize both a small bird and a large building in the same image.