In order to give consideration to the detection speed and accuracy of the defects of the nonmetallic adhesive structure of the rocket, a rocket defect detection algorithm based on the improved YOLOv5s X-ray image is proposed. Based on YOLOv5s, the algorithm uses deep separation convolution to redesign the Bottleneck structure in the feature extraction network, so as to improve the C3 module and improve the running speed by reducing the number of model parameters. Then the CBAM module is added after the Focus structure of the feature extraction network and before the convolution and upsampling of the Neck layer to improve the effective feature extraction of the model, make the model pay more attention to small targets, and try to maintain the running speed while improving the detection accuracy. The experimental results show that the mAP of the algorithm tested on the homemade rocket paste defect data set reaches 86.40%, which is 6.44% higher than the original model, and the FPS is 32 frames/second; Compared with SSD and YOLOX-Tiny network model, this model has excellent comprehensive performance in detection speed and detection accuracy, and can effectively detect the defects of non-metallic bonding structure of rocket.