<p> Edge cracks in rolled magnesium alloy plates directly affect yield rates and usable width. Currently, trimming allowances in cracked areas rely primarily on manual visual inspection and handheld measurement-based estimates, resulting in low efficiency, poor repeatability, and a high risk of over- or under-trimming. To address this, this paper integrates image processing techniques with deep convolutional neural networks. Based on the Seg-Net architecture and incorporating the cross-layer feature fusion mechanism from U-Net, we have developed an intelligent recognition model for edge crack detection and advanced quantitative feature analysis of magnesium alloy plates. To overcome the scarcity of magnesium alloy edge crack defect data, rolling experiments with varying process parameter combinations were designed to produce multiple edge-cracked defect plates. Crack images were collected to establish a specialized dataset. Using this dataset, DeepCrack, U-Net, and Seg-Net were trained for comparative evaluation. The results show that DeepCrack significantly outperforms U-Net and Seg-Net in terms of ODS, OIS, and AP metrics: compared to U-Net, it achieves improvements of 47.70%, 34.36%, and 21.71%, respectively, and compared to Seg-Net, it achieves improvements of 73.82%, 48.48%, and 69.04%, respectively. The results demonstrate that the DeepCrack model achieves high consistency with ground-truth annotations for key geometric features—including crack width, length, and orientation—when identifying edge cracks characterized by metallic luster, discontinuous fracture morphology, and uneven depth distribution. Compared to U-Net and Seg-Net, this model exhibits better performance. These results validate that the cross-layer fusion and multi-scale supervision mechanisms significantly enhance the accuracy of identifying surface cracks in rolled magnesium alloy plates.</p>