To ensure safe reactor operation, a variety of detection techniques are required to ensure the qualities of fuel pellets. To address high misdetection rate of cracks due to low contrast and complex background in the detection of surface cracks in fuel pellets, a surface crack detection algorithm based on convolutional neural networks (CNN) and the Beamlet algorithm is proposed. First, images are divided into equal-sized patches, which are used as training samples for the crack recognition model (CrackCNN). Then, the crack-containing region in the image is identified and located by the trained CrackCNN. Finally, a crack in identified region is detected by the Beamlet algorithm. The proposed method, which utilizes both CNNs and Beamlet, can improve detection accuracy and effectively reduce the probability of crack misdetection. Experimental results demonstrate that the F-measure of the proposed algorithm is enhanced by 6.4% and 3.4% compared to using only the Beamlet algorithm and using the double threshold and tensor voting algorithm, respectively.