An adaptive variational mode decomposition (AVMD) and recursive quantitative analysis technique (RQAT) for feature extraction was proposed combined with convolutional neural network (CNN) for recognition in order to address the issues of inaccuracy and unreliability in ultrasonic defect detection of carbon fiber reinforced plastics (CFRP). Six types of simulated defects were embedded in the experiments, and 500 A-scan waveform signals were collected for each defect type after ultrasonic phased array detection. The stingray intelligent optimization algorithm was used to optimize the K and Alpha values required for variational mode decomposition (VMD). Intrinsic mode function (IMF) components were obtained by using these optimized parameters, and high-frequency noise parts were filtered out. The remaining IMF components were processed with recursive quantitative analysis technique. Each signal yielded 72 feature values, which were assembled into feature vectors and input into the CNN for recognition. The recognition accuracy was 99.94% for the training set, 98.09% for the validation set, and 98.27% for the test set. Results show that the combination of AVMD, RQAT and CNN solves the defect recognition and classification problem in CFRP ultrasonic testing.