Abstract：With the increasing number of document forgery and economic dispute cases, the accurate identification of homochromy inks is of great significance in judicial expertise. Traditional methods (such as thin-layer chromatography and Raman spectroscopy) have limitations including sample destruction and time-consuming procedures. At the same time, hyperspectral imaging (HSI) has emerged as a promising alternative due to its advantages of image-spectrum integration and non-destructive detection. However, the existing ink classification methods based on “dimension reduction and clustering” are difficult to fully explore the nonlinear characteristics of high-dimensional data, and shallow machine learning models have limited expressive ability and are susceptible to information loss and error accumulation. Therefore, a deep learning model HI-CNN integrating multi-scale convolution and channel attention mechanism is proposed in this paper, and it is combined with hyperspectral imaging for the identification of homochromy inks. The model employs multi-branch parallel convolution to extract spectral features at different scales, comprehensively capturing spectral information across bands. And a channel attention mechanism dynamically enhances discriminative bands, focusing on key spectral information. Residual connection optimization gradient propagation is adopted to avoid gradient explosion and gradient vanishing, thereby reducing error accumulation and improving training efficiency. Experiments were conducted on the UWA Writing Ink Hyperspectral Image (WIHSI) dataset to determine the optimal training data partitioning and parameter settings. Ablation studies were designed to validate the effectiveness of the multi-branch parallel convolution structure, channel attention mechanism, and residual connections in improving model performance. Finally, the performance of the model proposed in this paper was compared with that of other model architectures on the current dataset. The experimental results show that the multi-branch structure and the channel attention mechanism improved the accuracy rates by 4.6% and 1.0% respectively, and the training cycle was shortened by 34% through the residual network connection. For the most challenging identification of the black ink, HI-CNN achieved an accuracy rate of 98.07% (an improvement of 5.3% compared to the optimal model CAE-LR). In comparison, for the identification of blue ink the accuracy rate reached 99.06%, which was generally superior to the existing methods. This study provided an accurate and efficient solution for identifying homochromy inks, thereby reducing reliance on professional expertise in forensic document examination. It had significant application value in the field of judicial expertise and promoted the leapfrog development of homochromy ink identification technology, transitioning from reliance on experience to scientific quantification.
姜林一，代雪晶，李云鹏，汤澄清. 基于改进卷积神经网络和高光谱成像技术的同色墨水鉴别[J]. 光谱学与光谱分析, 2025, 45(12): 3422-3430.
JIANG Lin-yi, DAI Xue-jing, LI Yun-peng, TANG Cheng-qing. Identification of Homochromy Inks Based on Improved Convolutional Neural Network and Hyperspectral Imaging Technology. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2025, 45(12): 3422-3430.
[1] Mariotti K C, Ortiz R S, Ferrão M F. Science & Justice, 2023, 63(3): 387.
[2] ZHOU Yi-han, CUI Lan, ZHANG Cheng, et al(周奕含, 崔 岚, 张 程, 等). Laser & Optoelectronics Progress(激光与光电子学进展), 2025, 62(16): 1.
[3] LI Shuo, CUI Lan, FU Pei(李 硕，崔 岚，付 沛). Chinese Journal of Inorganic Analytical Chemistry(中国无机分析化学), 2024, 14(6): 826.
[4] HAN Man-li, HOU Wei-min, SUN jing-guo, et al(韩嫚莉, 侯卫民, 孙靖国, 等). Journal of University of Electronic Science and Technology of China(电子科技大学学报), 2019, 48(1): 117.
[5] Devassy B M, George S. Forensic Science International, 2020, 311(6): 110194.
[6] WANG Shu-yue, YANG Yu-zhu, HE Wei-wen, et al(王书越, 杨玉柱, 何伟文, 等). Journal of Instrumental Analysis(分析测试学报), 2021, 40(10): 1489.
[7] CHEN Xin-gang, ZHANG Wen-xuan, MA Zhi-peng, et al(陈新岗, 张文轩, 马志鹏, 等). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2025, 45(4): 932.
[8] Mantripragada K, Dao P D, He Y, et al. PLoS One, 2022, 17(7): e0269174.
[9] López-Baldomero A B, Buzzelli M, Moronta-Montero F, et al. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 2025, 335: 125916.
[10] Khan Z, Shafait F, Mian A. Pattern Recognition, 2015, 48(11): 3615.
[11] Khan M J, Yousaf A, Abbas A, et al. Journal of Electronic Imaging, 2018, 27(5): 053001.
[12] Khan M J, Khurshid K, Shafait F. Proceedings of 2019 International Conference on Document Analysis and Recognition (ICDAR), 2019, 1097.
[13] Jaiswal G, Sharma A, Yadav S K. Computers and Electrical Engineering, 2022, 99: 107770.
[14] DENG Liang-jian, RAN Ran, WU Xiao, et al(邓良剑, 冉 燃, 吴 潇, 等). Journal of Image and Graphics(中国图象图形学报), 2023, 28(1): 17.
[15] CHEN Wen-hao, HE Jing, LIU Gang(陈文豪, 何 敬, 刘 刚). Laser & Optoelectronics Progress(激光与光电子学进展), 2022, 59(18): 1811001.
[16] BAO Cheng-xiao, JIANG Wei, WANG Yu-xiao(包丞啸, 姜 威, 王玉潇). Optical Technology(光学技术), 2019, 45(6): 756.
[17] Luo Zhipei, Shafait F, Mian A. Proceedings of 2015 13th International Conference on Document Analysis and Recognition (ICDAR), 2015, 496.
[18] Abbas A, Khurshid K, Shafait F. Proceedings of 2017 14th International Conference on Document Analysis and Recognition (ICDAR), 2017, 1229.
