The measurement of crack length is fundamental to the evaluation of crack risk and a prerequisite for crack repair. Aiming at the problems of irregular shapes, small target, sparse data sets and distortions of crack imaging angle, a sparse data driven method was proposed to extract the surface crack length of turbine blades. Firstly, to enhance the Unet model′s precision in handling sparse data, we employ a combination of the GeLu function with the Vgg16 network for feature extraction. The extracted features then serve as inputs for the Unet network′s decoding process. To ensure model compatibility, we incorporate pre-trained weights into the randomly initialized weights and integrate an efficient pyramid compression attention module into the skip connection layer. This approach significantly improves the model′s capability to focus on crack characteristics amidst complex backgrounds. Then, in order to get the unit pixel characteristic curve of the crack, after the fine segmentation, a skeleton structure with eight neighborhood is proposed to preserve the crack backbone characteristic structure. Finally, through an in-depth analysis of camera imaging principles, we discuss the impact of blade chord angles and camera parameters on crack length measurements, establishing a conversion model between pixel size and actual dimensions. Experimental results indicate that when the measuring distance ranges from 100 to 300 mm, the maximum error in crack length is 6. 8% . Compared to X-ray measurements, our method proves to be a viable alternative for measuring the surface crack length of turbine blades. Moreover, the enhanced algorithm demonstrates greater accuracy than the original algorithm in detecting sparse data, with an average cross-over ratio improvement of 7. 14% . The proposed method offers a theoretical foundation and data support for evaluating blade quality and guiding subsequent repairs.