Carbon fiber composites are widely used in different fields owing to their unique properties. However, the presence of defects adversely affects the performance of the material, thus causing significant economic losses and safety hazards. In this study, a transmissed-terahertz continuous wave detection system was constructed to detect samples with internal defects, perform image preprocessing on the detection results, and construct the target dataset. To address industrial scenarios involving significant image background interference, easy confusion of defect categories, significant variation of defect scales, and unsatisfactory detection of small defects, an improved YOLOv5s algorithm was proposed, which improves the accuracy and speed of intelligent defect recognition by increasing the small-target detection layer and the attention network's convolutional block attention module (CBAM), as well as by improving the loss function. Experimental results show that the improved YOLOv5s algorithm yields an accuracy of 92.3% and a recall of 80.8% on the test set, which are 10.6 percentage points and 3.0 percentage points higher than those of the original YOLOv5s algorithm, respectively. Furthermore, it exhibits better feature extraction and greater robustness, which eliminates the issue of misdetection.