In continuous terahertz imaging systems, the low intensity of the terahertz source and susceptibility to external environmental interference result in issues such as low signal-to-noise ratio and blurred boundaries in the images. To address these problems, a method utilizing the Unet deep learning network for image defect recognition is proposed. Theoretically, a training dataset is constructed using terahertz images collected through experiments, and the Unet network is trained with this dataset to learn the feature information of the data. Subsequently, the network predicts and outputs defects in images from the test dataset. An experimental system is built concurrently with the theoretical design for validation, and quantitative analysis is conducted by calculating the information entropy and EOG values of the images before and after recognition. The results indicate that the predicted images exhibit an 8.03% increase in information entropy and a 13.61% increase in EOG values compared to the original images, demonstrating that the network significantly enhances image clarity and visual quality.