For the safe operations of nuclear power plants, it is important to inspect the reactor internal components frequently. However, current practice involves human technicians who review the inspection videos and identify cracks on metallic surfaces of underwater components, which is costly, time-consuming, and subjective. Detecting cracks on metallic surfaces from the inspection videos is challenging since the cracks are tiny and surrounded by noisy patterns in the background. While other crack detection approaches require longer processing time, this study proposes a new approach called NB-fully convolutional network (NB-FCN) that detects cracks from inspection videos in real time with high precision. An architecture design principle is introduced for FCN, where the FCN can take image patches for training without pixel-level labels. Based on the naïve Bayes (NB) probability, a parametric data fusion scheme called pNB-Fusion is proposed to fuse crack score maps from multiple video frames and outperforms other fusion schemes. The proposed NB-FCN achieves 98.6% detection average precision (AP) and requires only 0.017 s for a 720 × 540 frame and 0.1 s for a 1920 × 1080 frame. Based on its capability and efficiency, the proposed NB-FCN is a significant step toward real-time video processing for autonomous nuclear power plant inspection.