The aircraft flat-tail honeycomb sandwich composite plate has complex structure, large area and many types of defects. The bonding state between honeycomb and skin can be intuitively analyzed by ultrasonic C-scan imaging. As a result, a large number of detection images need to be evaluated by the rich work experience of technicians, there are problems such as low evaluation efficiency and strong subjectivity. Therefore, an intelligent classification technology of ultrasonic C-scan image of honeycomb sandwich composite plate bonding layer based on deep learning is proposed. Firstly, an ultrasonic C-scan testing image of the interface between the adhesive layer and the honeycomb is acquired by an interface reflected wave tracking method, and the image quality is further improved by combining an image processing technology. Secondly, in order to construct the training database, C-scan image samples are intercepted by sliding window, and the samples are divided into three bonding states (target areas) according to the amplitude distribution of C-scan, and a small sample image data set expansion method is proposed. Finally, the 50-layer residual network (ResNet50) is constructed and trained, and the classification ability of the deep learning network for honeycomb bonding states is evaluated. The results show that the interface reflection wave tracking can overcome the shape change of the skin surface and form the C-scan image of the honeycomb bonding layer, and the ResNet50 network can identify the three types of target areas of the honeycomb sandwich composite panel structure with good stability and accuracy, and reflect the characteristics of “intelligence”.