Abstract A method is reported using attenuated total reflection – Fourier transform infrared (ATR-FTIR) and chemometrics analysis for the forensic discrimination of ship deck paint. The automatic baseline correction, peak area normalization, multiple scattering correction and Savitzky-Golay algorithm using smoothing were adopted to preprocess the spectral data. Several pattern recognition methods including principal component analysis (PCA), Fisher discriminant analysis (FDA), and K-nearest neighbor analysis (KNN) were adopted as the algorithms for constructing classifiers. The results showed that in the principal component analysis model, the scores of 5 brands of samples were different from each other. The derivative spectroscopy revealed hidden differences in the original spectra with improved resolution. In the Fisher discriminant analysis model, samples achieved a more ideal discrimination result. In K-nearest neighbor analysis model, 1 was selected to be the optimal K value to construct the classification model and the discrimination result was ideal. Fisher discriminant analysis was better than principal component analysis and the K-nearest neighbor analysis in the ability to discriminant between samples. It is important to use multiple indicators to evaluate and assess the classification results instead of a single indicator. The precision rate, recall rate, and F-measure may be considered except for the total accuracy in evaluation and assessment.

中文翻译：

---

使用衰减全反射 - 傅里叶变换红外 (ATR-FTIR) 光谱和化学计量分析法无损识别船舶甲板涂料

摘要 报告了一种使用衰减全反射 - 傅里叶变换红外 (ATR-FTIR) 和化学计量学分析对船舶甲板油漆进行法医鉴别的方法。采用自动基线校正、峰面积归一化、多重散射校正和使用平滑的Savitzky-Golay算法对光谱数据进行预处理。采用主成分分析（PCA）、Fisher判别分析（FDA）和K-最近邻分析（KNN）等几种模式识别方法作为构建分类器的算法。结果表明，在主成分分析模型中，5个品牌的样本得分不同。衍生光谱揭示了原始光谱中隐藏的差异，并提高了分辨率。在Fisher判别分析模型中，样本取得了较为理想的判别结果。在K-近邻分析模型中，选取1作为构建分类模型的最优K值，判别结果较为理想。Fisher判别分析在判别样本的能力上优于主成分分析和K-近邻分析。重要的是使用多个指标而不是单个指标来评估和评估分类结果。除了评估和评估的总准确率外，可以考虑准确率、召回率和 F-measure。选择1为最优K值构建分类模型，判别结果理想。Fisher判别分析在判别样本的能力上优于主成分分析和K-近邻分析。重要的是使用多个指标而不是单个指标来评估和评估分类结果。除了评估和评估的总准确率外，可以考虑准确率、召回率和 F-measure。选择1为最优K值构建分类模型，判别结果理想。Fisher判别分析在判别样本的能力上优于主成分分析和K-近邻分析。重要的是使用多个指标而不是单个指标来评估和评估分类结果。除了评估和评估的总准确率外，可以考虑准确率、召回率和 F-measure。