【作者】 张浩；

【导师】 高树辉；

【作者基本信息】 中国人民公安大学 ， 刑事科学技术， 2024， 硕士

【摘要】 交通肇事逃逸案件中,肇事者往往有意识或无意识地毁灭或消除现场痕迹物证,增加了案件侦破难度。在这些案件中,车身油漆碎片是最为常见的微量物证,也是现场勘查人员最为关注的证据之一,它有助于查找、追踪涉案车辆,有效缩小侦查范围,能为案件的侦破提供有效支持。汽车车身油漆物证的传统仪器分析检测方法较为成熟,多在实验室内完成,但样品前处理繁琐,不能完全满足现场物证原位快速无损检测需求。为了解决上述瓶颈问题,本文基于高光谱成像技术结合模式识别方法,通过深度挖掘车漆物证的高光谱数据,建立了快速、准确且无损的车漆物证鉴别模型,为案件中汽车车身油漆物证快速智能检验提供了新的技术支持和方法补充。主要内容包括:(1)构建了丰富汽车车身油漆物证高光谱数据集。采集了8种颜色类别的来自30种车型的油漆样本108份。每个样本高光谱图像选取固定感兴趣区域(ROI),对区域内10×10像素融合后的光谱作为数据集,每个样本ROI为365个光谱数据,30种车型108个车漆样本共获得39420个光谱数据,同时研究颜色、时间、原漆与修补漆以及车型对车漆平均光谱的影响。为模型的训练和验证提供了丰富的数据支持,大大提高了分类的准确性和稳定性。(2)汽车车身油漆物证高光谱数据挖掘及机器学习实验研究。针对白色原车漆光谱数据,运用传统机器学习算法组合进行分类,基于S-G卷积平滑、导数处理和多元散射校正,提升数据质量并减少噪声影响。通过SPA和CARS算法进行降维,构建了基于SVM、RF和LR的模型,实验结果证明,D1-CARS-SVM、S-G-SPA-RF和MSC-SPA-LR分别在各自模型中实现了最高准确率,分别为93.60%、93.23%和91.53%,其中D1-CARS-SVM模型的表现最为优异,证明了高光谱成像技术结合机器学习算法在汽车车身油漆物证识别中应用潜力。(3)创新性地改进并构建了的一维深度残差收缩网络(1D-DRSN)。通过使用一维卷积层替换二维卷积层,采用全局最大池化的注意力机制,并融入Dropout层,构建了自定义网络。实验结果表明,1D-DRSN模型在训练集和测试集上的准确率分别为99.51%和98.60%,损失函数值分别下降到0.093和0.106收敛。该模型的分类效果和精度与一维卷积神经网络(1D-CNN)模型和三种传统的机器学习方法支持向量机(SVM)、随机森林(RF)、逻辑回归(LR)相比,分类性能显著提高,实现了高光谱技术的汽车车身油漆物证智能识别。(4)在1D-DRSN基础上,开发了一款车漆物证碎片的高光谱数据分析软件,软件可展示载入的高光谱数据图像,高效处理和分类光谱数据信息,或者训练自己的模型。最终分类结果能在界面的“Outcome”部分清晰显示,从而大幅提升工作效率和判别的准确性。综上,本文打破了车漆物证的传统检验方法,创新性地结合模式识别并自行设计搭建了汽车车漆物证识别模型和检测软件,在自建的数据集上取得了优秀的分类效果,建立了光谱数据自动分类模式,利用模拟测试验证了本文方法在物证光谱采集后可实现快速自动分类,实现了现场油漆物证原位快速无损检测需求。更多还原

【Abstract】 In hit-and-run cases,the perpetrator often consciously or unconsciously destroy or eliminate the scene traces of physical evidence,increasing the difficulty of case detection.In these cases,body paint fragments are the most common trace physical evidence,but also one of the most concerned about the evidence of the scene investigators,it helps to find and track the vehicle involved in the case,effectively narrowing the scope of investigation,can provide effective support for the detection of the case.The traditional instrumental analysis and detection method of auto body paint evidence is relatively mature,mostly completed in the laboratory,but the sample pretreatment is cumbersome,and cannot fully meet the demand for rapid non-destructive testing of on-site physical evidence in situ.In order to solve the above bottlenecks,this paper is based on hyperspectral imaging technology combined with pattern recognition methods,through the depth of mining the hyperspectral data of the paint exhibits,to establish a fast,accurate and non-destructive identification model of the paint exhibits,which provides new technical support and methodological supplements for the rapid and intelligent inspection of the auto body paint exhibits in the case.The main contents include:(1)A rich hyperspectral dataset of auto body paint exhibits was constructed.108 paint samples from 30 car models in 8 colour categories were collected.A fixed region of interest(ROI)was selected for each sample hyperspectral image,and the fused spectra of 10 × 10 pixels in the region were used as the dataset.365 spectral data were obtained for each sample ROI,and a total of 39,420 spectral data were obtained from 108 paint samples of 30 car models,and the effects of colour,time,original and refinish paints,and car models on the average spectra of car body paints were investigated at the same time.Rich data support is provided for the training and validation of the model,which greatly improves the accuracy and stability of classification.(2)Experimental research on hyperspectral data mining and machine learning for auto body paint physical evidence.For the white original car paint spectral data,a combination of traditional machine learning algorithms is used for classification,based on S-G convolutional smoothing,derivative processing and multivariate scattering correction,to improve the data quality and reduce the impact of noise.The models based on SVM,RF and LR were constructed by dimensionality reduction with SPA and CARS algorithms,and the experimental results proved that D1-CARS-SVM,S-G-SPA-RF and MSC-SPA-LR achieved the highest accuracy in their respective models,which were 93.60%,93.23% and 91.53%,respectively,with the D1-CARS-SVM model has the best performance,proving the potential of hyperspectral imaging technology combined with machine learning algorithms in the application of auto body paint object recognition.(3)One-dimensional deep residual shrinkage network(1D-DRSN)is innovatively improved and constructed.The custom network was constructed by using a 1D convolutional layer instead of a 2D convolutional layer,adopting the attention mechanism of global maximum pooling,and incorporating a Dropout layer.The experimental results show that the accuracy of the 1D-DRSN model is 99.51% and 98.60% on the training and test sets,respectively,and the loss function values drop to 0.093 and 0.106 convergence,respectively.The classification effect and accuracy of the model are significantly improved compared with the one-dimensional convolutional neural network(1D-CNN)model and three traditional machine learning methods support vector machine(SVM),random forest(RF),and logistic regression(LR),which achieves intelligent identification of automobile body paint exhibits with hyperspectral technology.(4)On the basis of 1D-DRSN,a hyperspectral data analysis software for car paint evidence fragments was developed,which can display the loaded hyperspectral data images,process and classify the spectral data information efficiently,or train its own model.The final classification results can be clearly displayed in the "Outcome" part of the interface,which can greatly improve the working efficiency and accuracy of the judgement.更多还原