This paper takes visual communication design as the entry point and reduces the complexity of the network model and the number of parameters by a convolutional neural network. The Faster R-CNN algorithm has been improved and is now used as the main algorithm for traditional element extraction and detection. The global detection and segmentation extraction algorithm based on U2-Net is utilized to reduce the depth of the network model, and the VGG16 algorithm is improved according to the characteristics of traditional elements, migration learning and its network structure in order to increase the sensory field and solve the problem of loss of accuracy of the feature maps on the basis of not changing the feature layer. An empirical study of traditional element extraction is conducted by using shadow costume patterns as an example. The results show that in multiple shadow images, the range of hues used for the five character types, namely, small Dan, Sheng, Jingsheng, Chou and old Dan, is