<p>Crack detection is a vital step in the quality assessment of metal products. Conventional image processing techniques often use several dominant features to detect cracks but tend to produce numerous false positives, especially in panels with complex geometries. Although deep learning-based methods demonstrate high performance, they typically require large and balanced datasets. When crack data is significantly smaller than healthy data, models tend to identify all inputs as healthy This study presented a two-stage strategy to achieve robust crack detection performance on large metal panels with limited data. To partially address the issue of limited crack data and data imbalance, we optimized an edge line analysis method for extracting crack candidates. The dominant feature of cracks, characterized by their sharpness, is leveraged to extract candidates. These candidates isolate crack-like regions, thereby reducing the model’s exposure to irrelevant background and non-defect areas. This targeted focus contributes to a more balanced and efficient dataset for training and inference. With balanced and targeted datasets, deep learning models can more effectively achieve high performance. We fine-tuned a pre-trained deep neural network using these candidates to assess the likelihood of cracks in each input. We conducted comprehensive experiments using images captured from large stamped metal panels in an environment resembling real manufacturing settings. Notably, the only required devices are a webcam and a standard desktop. The results showed that our proposed technique significantly reduced false positives and effectively enhanced crack detection performance across various datasets, indicating the potential for application in real-world industries.</p>