Abstract: Micro-cracks generated during ultrasonic welding of thin sheet metal seriously affect the welding quality. By taking Cu/Al ultrasonic welded joints as objects, an improved CenterNet network was proposed and applied to ultrasonic welding crack detection. An improved network was formed by an anchor-free CenterNet network and an FFM module and was embedded into the entire process of training and detection. A dataset and template samples were generated through image processing technology combined with data augmentation. A FLANN-accelerated ORB template matching algorithm was adopted to locate the surface welding region. The feature fusion module of the improved network was utilized to complete the detection of fine welding cracks. Finally, the welding crack detection effect was verified by comparing the detection indicators of the feature fusion ablation experiment and the improved network. The results indicate that the accuracy of the adopted template matching algorithm in locating the welding head indentation region reaches 95%. The accuracy of the improved CenterNet network for Cu/Al ultrasonic welding crack detection is 93%, and the detection speed reaches 33.13 frames/s, which are increased by 24.0% and 63.2% respectively compared with the ablation experiment. The total parameter quantity of the crack detection network is 3.04 M, which classifies it as a lightweight network.