<p>Crack detection is of great importance in industrial production. However, the existing crack detection methods cannot always meet the needs of detection. When dealing with cracks with complex shapes, the existing crack detection methods can only recognize them roughly, which brings great inconvenience to the subsequent repair work. When dealing with tiny cracks, existing methods often fail to recognize them, resulting in missed detections. When dealing with cracks that have complex backgrounds, factors such as zebra stripes, oil stains, and other background interferences may lead to false detection, resulting in errors in crack identification. To solve these problems, we design a network specifically to deal with these unconventional cracks. This network first locates the cracks by hierarchical searching, then optimizes the crack contour by feedback iterations of the features, and finally enhances the detailed information of the cracks. Meanwhile, we also collect a large number of unconventional crack samples and construct an unconventional crack dataset. On this dataset, the proposed network outperforms the existing SOTA segmentation network in several evaluation indexes and achieves the detection of unconventional cracks. Access to the dataset and relevant code is provided through: https://github.com/U-CRACK/FOUCNet.</p>