As an important component of nuclear reactors, guide tubes have critical applications in the nuclear industry. In order to assist engineers in quickly and accurately performing guide tube defect detection tasks and to construct an automatic defect detection system for nuclear components, this study proposes an automatic detection method that integrates traditional image processing algorithms with U-Net and YOLOv8 deep learning algorithms. This method is aimed at the data collected by metallographic microscopes, using the U-Net network to extract edge coordinate information of the data, and designing a boundary fitting method based on smooth optimization to obtain the boundary equation of the fusion zone. At the same time, the YOLOv8 network is used to detect the coordinate range of weld joints, and the final coordinates of the weld joints are obtained by combining the coordinate of the fusion zone edge. Experimental results show that, while maintaining fitting accuracy, the boundary equation obtained by the smooth optimization algorithm has better continuity. Compared with the traditional method of metallographic inspection combined with manual measurement, this method significantly improves the detection efficiency and accuracy of Effective Melt Depth and Melt Zone Depth of guide tubes.