It is important for the accurate identification of cracks of engineered cementitious composites (ECC) to investigate the mechanical properties and durability of ECC. To solve the problems like the large number and density of ECC cracks and heavy noise interference, this study adopted the U-NET model suitable for biological image recognition, and optimized part of ResNet network layer structure based on the deep learning method. This study also used the neural network model and combined with the created data suitable for the ECC, and performed the semantic segmentation to obtain the crack pixels. For crack parameter extraction, this study used the bone extraction method and combined with digital image processing process, and used the CLAHE filter and half-peak full-width concept to obtain the crack width. The crack identification and parameter extraction method was applied to detect the cracks on hybrid-fiber ECC dog bone specimen and ECC link slab. The results show that the error range between the ECC crack identification and intelligent detection established by the deep learning method and the actual manual measurement is within 0.6 mm. The results of this study can provide an accurate, effective and high-throughput analysis for the ECC crack inspection and feature quantitative identification.