Structural crack detection is of paramount importance for ensuring the safety and long-term integrity of critical infrastructure. Conventional crack detection methodologies often prove inadequate in preserving the continuity of fine crack edges and effectively handling the complexities of multi-directional crack propagation. To overcome these limitations, we introduce EMCrossNet, a novel crack detection architecture to enhance crack edge continuity and robustly capture multi-directional propagation patterns. EMCrossNet is composed of three core module, EdgeSculptor, which leverages custom-designed vertical and horizontal convolutional operators to meticulously enhance subtle edge details; BiFusion, which dynamically integrates multi-directional features through an innovative bi-directional scan-adaptive fusion mechanism; and ScaleAdaptive Segmentor, which achieves accurate crack segmentation across multiple scales via hierarchical dilated filtering and spatial feature recalibration techniques. Comprehensive experimental evaluations conducted on diverse multi-scene benchmark datasets reveal that EMCrossNet significantly surpasses state-of-the-art methodologies, achieving a superior F1 score of 0.8147 and a markedly improved mIoU of 0.8287.