文献类型：期刊文献

英文题名：Astragalus Fingerprint Identification Based on Generative Adversarial Network and Encoder-Decoder

作者：Li, Sihai[1]; Liu, Mingqi[1]; Song, Hang[1]

第一作者：李四海

机构：[1] College of Information Engineering, Gansu University of Chinese Medicine, Gansu, Lanzhou, 730000, China

第一机构：甘肃中医药大学

年份：2025

外文期刊名：SSRN

收录：EI(收录号：20250143409)

语种：英文

外文关键词：Benchmarking - Deep neural networks - Financial data processing - Generative adversarial networks - Image retrieval - Information management - Network coding - Network security - Palmprint recognition - Photomapping - Polycyclic aromatic hydrocarbons - Steganography

摘要：Astragalus has a subtle odor and a slightly sweet taste, which possesses medicinal properties by tonifying qi and elevating yang, fixing the epidermis, stopping sweating, inducing diuresis, and reducing swelling. Due to variations in origin, moisture content, and other factors, there are noticeable differences in appearance, leading to the presence of counterfeit products in trial production. The traceability of Astragalus' origin can better address this issue. Currently, deep convolutional networks are widely used in fingerprint mapping data analysis. In contrast, fingerprint mapping data features tend to be more. The convolutional neural network mainly focuses on the local relationship between features, and it is difficult to capture the global relationship between features. In contrast, fingerprint mapping collection tends to consume a large amount of workforce, material, and financial resources, and the amount of data tends to be less, or the sample category is seriously unbalanced. Therefore, this paper proposes a novel DCNN model. It addresses the issue of limited data by enhancing Astragalus fingerprint atlas data through a CTGAN network. Furthermore, an encoder-decoder structure is integrated into the 1D-CNN network, enabling the simultaneous capture of global and local feature relationships, thereby improving classification performance. The experimental results demonstrate that, after data augmentation via the CTGAN network, the accuracy of the four models (PLS-DA, SVM, KNN, and 1D-CNN) is 97.3%, 95.3%, 96%, and 97.3%, respectively. This represents improvements of 1.7%, 6.4%, 4.9%, and 1.7% compared to the original data. Additionally, the 1D-CNN network with an encoder-decoder structure achieves an accuracy of 97.8% on the original data, surpassing the four models by 2.2%, 8.9%, 6.7%, and 2.2%, respectively. Finally, the proposed DCNN model, comprising CTGAN, 1D-CNN, and encoder-decoder networks, achieves an accuracy, recall, and F1 score of 98.67%, 98.72%, and 98.68% on Astragalus fingerprint maps. The accuracy metrics surpass the four models by 3.1%, 3.1%,9.8%, and 7.6%, respectively. Additionally, this study evaluates the model on a publicly available fruit puree dataset, where the DCNN model demonstrates 4%, 1.4%, 3%, and 8.5% higher accuracy than the four models, respectively. These results demonstrate the DCNN model's superior classification performance and generalization ability, offering a novel approach for fingerprint mapping research. ? 2025, The Authors. All rights reserved.