In order to realize the intelligent identification of internal lamination defects of aviation glass fiber composites, a spectroscopy system with multidegree of freedom fiber coupling terahertz time domain is built. The samples with simulated internal lamination defects are detected, and the detection results are filtered, enhanced and marked, and the data sets for target detection are constructed. At the same time, a modified YOLOv4 algorithm is proposed to improve the accuracy of intelligent defect recognition. The experimental results show that the improved YOLOv4 algorithm achieves 91.05% accuracy and 92.02% recall rate in the test set, which is 5.73% and 8.51% higher than the original YOLOv4 algorithm, respectively. This algorithm is characterized by a stronger feature extraction capability and good robustness, and obviously eliminates the error detection and omissions of the original YOLOv4 algorithm.