The aircraft engine blades working in extreme conditions may develop cracks, posing serious safety risks. The performance of deep learning methods is poor in engine blade crack detection due to limited number of crack surface images and small sizes. Regarding this, this paper introduces a small-sample engine blade surface crack detection method FCU2C, which integrates U2-Net and Canny. Firstly, the U2-Net is enhanced with the integration of the coordinate attention mechanism for better detection precision, called CA-U2Net. Secondly, CA-U2Net is used as feed-forward of FCU2C to improve noise resistance and robustness. The Canny operator segments the feed-forward outputs. Finally, the proposed method is trained on dataset CFD and transferred to predict aircraft engine blade cracks. Experiments demonstrate that our technique enhances the F1-score by 55.65%, 28.98%, and 9.94% when benchmarked against deeplabV3+, U-Net, and U2-Net, respectively.