Because of the demand for nondestructive testing of the bonding quality of soft foam composite structures, excitation mode research was conducted, and the pulsed infrared thermal image mode was determined suitable. A first-derivative processing method was used to form a featured image, which improved the image signal-to-noise ratio and defect resolution. Debonding defects in the specimens were effectively identified using the constructed neural network recognition software. The results revealed that the risk of false and missing detections in manual evaluation can be reduced by intelligent detection, and that the detection efficiency can be improved. The above research meets the requirements for the rapid detection of soft foam composite structures in spacecrafts.