融合HSI的深度残差收缩网络鉴别变造文件字迹油墨研究

Research on Combining Hyperspectral Imaging and Deep Residual Shrinkage Network for Ink Identification of Question Document Handwriting

经济犯罪和各类民事纠纷等案件中,字迹油墨检验对质疑文书同一认定有重要意义,相关研究一直是法庭科学安全领域的重要课题。鉴于传统方法效率和精度较低,提出一种结合高光谱图像的深度残差收缩网络快速且无损鉴别字迹油墨种类的新方法。首先,采集了30支不同品牌型号的黑色签字笔油墨的高光谱图像,对每支中性油墨的高光谱图像进行分割,提取笔迹区域进行10×10像素融合,获取了共计13942像素点的反射率数据作为样本集。其次,结合残差网络、软阀值化与注意力机制,提出适用于处理高光谱数据的一维深度残差收缩网络模型,同时将其与卷积神经网络和传统机器学习模型进行比较。实验得出,支持向量机、逻辑回归、随机森林3个模型的像素反射率值测试准确率分别为59.1%、57.8%和51.7%,卷积神经网络为64.2%,损失函数值下降到1.536,而深度残差收缩网络验证识别率最高,达到75.4%,损失函数值最终下降到0.920,达到收敛。实验结果表明,光谱检测方法具有无损、成像快速、操作简单的优点,提出的光谱检测深度残差收缩网络模型在笔迹油墨的分类效果和精度上具有明显优势,可实现黑色签字笔油墨种类高光谱数据的深度挖掘和准确分类,结合笔迹检验技术能为法庭科学中质疑文书检验提供技术支撑。

In the cases of economic crimes and various civil disputes,handwriting ink identification is of great significance for consensus identification in questioned documents,and the related research has been an important topic in the field of court science.In view of the low efficiency and accuracy of traditional methods,a new method combining hyperspectral imaging with deep residual shrinkage network was proposed for rapid and non-destructive identification of ink types.First,the hyperspectral images of 30 black signature pen inks of different brands and models were collected.The hyperspectral images of each neutral pen ink were segmented,and the handwriting regions were extracted for 10×10 pixel fusion to obtain a total of 13942 pixel points of reflectance data as the sample set.Second,a one-dimensional deep residual shrinkage network model suitable for processing hyperspectral data was proposed by combining residual network,soft valorization and attention mechanism.Meanwhile,this model was compared with convolutional neural network and traditional machine learning models.The experimental results showed that the test accuracy of pixel reflectance of support vector machine,logistic regression and random forest was 59.1%,57.8%and 51.7%,respectively,and that of convolutional neural network was 64.2%.The value of loss function dropped to 1.536.While the validation recognition rate of deep residual shrinkage network was the highest,reaching 75.4%,and the value of loss function finally dropped to 0.920,reaching convergence.The results showed that the spectral detection method had the advantage of non-destructive,fast imaging and simple operation.The proposed deep residual shrinkage network has obvious advantages in classification effect and accuracy of handwriting inks and can realize the deep mining of hyperspectral data and accurate classification of black signature pen ink types,which provides a technical support for the identification of questioned document in forensic science.