Abstract: To achieve intelligent recognition of internal metallic foreign objects in ultrasonic C-scan images of pouch cells and alleviate the impact of scarce ultrasonic C-scan images on image recognition models, this study proposes an image generation method that combines a small number of ultrasonic C-scan images with data-driven strategies. By analyzing the relationship between the pixels of ultrasonic C-scan images and the internal defects and structures of pouch cells, 300 samples are generated. An intelligent detection model is constructed using an improved U-Net architecture and evaluated using 10-fold cross-validation. The generalization capability of the proposed model is verified based on newly fabricated pouch cells containing copper particles with a particle size of 0.2–0.4 mm. Results show that the loss function decreases continuously, and the validation loss drops in tandem, with a reasonable gap relative to the training loss. The study shows that the model achieves a Dice coefficient exceeding 0.95 and an Intersection over Union (IoU) above 0.90 for image segmentation. The proposed method can effectively mitigate the problem of sample scarcity and successfully detect all copper particles within the 0.2–0.4 mm range, thereby providing a solution for the intelligent recognition of internal metallic foreign objects in pouch cells.