Abstract: The geometry and location of delamination defects inside a laminated composite insulating structure can cause field distortion during operation, leading to partial discharges and even insulation breakdown, so it is important to accurately detect the degree of delamination inside a laminated composite insulating part. Firstly, the terahertz spectroscopy is used to image the laminated composite insulation structure with delamination defects in frequency domain and a typical image set of delamination defects is obtained. Based on this, the DCGAN model is used to expand the images and a data set is built to achieve sample expansion and equalization. Finally, the geometric area of defects in the delamination region of the defect samples is calculated and analyzed by three SSA-CNN(semantic self-attentive) models. The results show that the DeepLabV3+(MobileNetV2) model has the highest pixel accuracy and the recognition rate of the layered region can reach 97.59%, and the geometric size of the defects in the layered region can be successfully characterized by the calculation of pixel points. The results of the study can provide a technical reference for the non-contact quantitative characterization of delamination defects inside laminated composite insulation structures.