Abstract: Micro-cracks generated during ultrasonic welding of thin sheet metal significantly compromise weld quality. Focusing on Cu/Al ultrasonic welded joints, this study proposes an improved CenterNet for micro-crack detection in ultrasonic welds. An anchor-free CenterNet network incorporating Feature Fusion Modules (FFM) was developed, enabling seamless integration throughout the training and detection pipeline. A dataset and template samples were generated using image processing techniques combined with data augmentation. The FLANN-accelerated ORB template matching algorithm was first employed to locate the surface weld region. Subsequently, the feature fusion capabilities of the improved network were leveraged to detect fine welding cracks. An ablation study on feature fusion was conducted to validate the crack detection effectiveness. Results demonstrate that the adopted template matching algorithm achieves 95% accuracy in locating the sonotrode indentation zone. Compared to the ablation study, the improved CenterNet attained a crack detection accuracy of 93% for Cu/Al ultrasonic welding, representing an improvement of 24.0%. The detection speed reached 33.13 frames per second (fps), an increase of 63.2%. The total parameter count of the crack detection network is 3.04 million, classifying it as a lightweight network.