Traditional bullet trace detection generally uses laser to detect rifling traces to extract the signal of the rifling traces. The signal form is complex and random, which severely restricts its practical application. Aiming at the problems of low precision and complicated operation of traditional bullet trace detection, this paper adopts multi-scale registration, elastic shape measurement and convolutional neural network technology, and uses adaptive control method based on multi-mode elastic drive to establish the end of the specimen. Position and attitude parameter distribution model. At the same time, the isolated forest algorithm is used to detect the signal for anomaly processing, and the variable-scale morphological filtering algorithm is used to remove non-small features. The square velocity function is introduced to optimize the elastic shape measurement algorithm to complete the curve contour embedding layer mapping. Aiming at the matching part of the rifle line shape, a convolutional neural network model of optimized parameter sharing connection triples suitable for trace features is established, and the network is trained to convergence by calculating the similarity of the embedding layer and minimizing the triple loss function. The method solves the accuracy and operability problems faced in the traditional bullet trace detection, the cost is greatly reduced compared with the traditional detection method, and the stability of the detection result can be guaranteed. The comparison of similarity matching experiment results can also strongly prove the superiority of this method.