The defect detection is an important task for maintaining the hydro-junction project. A two-stage crack defect segmentation method based on target detection network is proposed to solve the problem of severe brightness imbalance and large noise in dam surface images. In the first stage, to improve the ability to locate crack areas, Inception Resnet V2 is used as feature extraction network to help Faster-RCNN extract more effective deep features, and the brightness, contrast of image is randomly adjusted before training. In the second segmentation stage, the crack areas are segmented at pixel-level using K-means. The experimental results on the self-made crack image dataset show that the location accuracy (AP) of the crack areas can be improved by 1.9%, reaching 96.8%, compared with other segmentation networks that do not locate crack areas, the intersection over union for segmentation of cracks (lou) of the final segmentation results is at least 9.4% higher, reaching 52.7%. This method can provide effective technical support for inspection work of hydro-junction project.

缺陷检测是维护水电枢纽工程的重要任务。提出了一种基于目标检测网络的两阶段裂纹缺陷分割方法，以解决大坝面图像亮度严重不均衡，噪声大的问题。在第一阶段，为了提高定位裂纹区域的能力，将Inception Resnet V2用作特征提取网络，以帮助Faster-RCNN提取更有效的深层特征，并在训练之前随机调整图像的亮度，对比度。在第二分割阶段，使用K-均值在像素级别分割裂纹区域。自制裂纹图像数据集的实验结果表明，定位精度（AP）的裂纹区域可提高1.9％，达到96.8％，与未定位裂纹区域的其他分段网络相比，最终分段结果的裂纹联合的交集（lou）至少高9.4％ ，达到52.7％。该方法可为水电枢纽工程的验收工作提供有效的技术支持。