Crime Scene Sketches Classification Based On Convolutional Neural Network

Crime scene sketch,as an important part of the investigation records,plays an important role in the reconstruction of crime scene and criminal investigation.Nowadays,with the booming development of Big Data and Artificial Intelligence,using computer-assisted for analyzing crime scene sketches will become a growing trend of ‘Informational Investigation '.However,Automated method of classifying crime scene sketches is the basis of achieving that goal.This paper raised a challenge of classifying crime scene sketches base on the practical necessity,and proposed an automated method for classifying crime scene sketches based on Convolutional Neural Network.Firstly,by consulting relevant materials,we summarized a variety of classifying standards and drawing standards.Based on that,this paper proposed a standard for drawing and classifying sketches used in this study.Secondly,the types of crime scene sketches should be covered in the crime scene sketch dataset,via consulting experienced staff and gathering information from the National Criminal Scene Investigation Information System(NCSIIS).Then we successively built the basic version and improved version of crime scene sketch dataset.There are 64 098 crime scene sketches and 27 162 photos in the improved version dataset.Thirdly,based on the classic Convolutional Neural Network architectures,we successively designed a new CNN called CSSNet(as ‘Crime Scene Sketch Net')for the basic crime scene sketch dataset and a novel CNN named XCTNet(as ‘XianChangTu Net')for the improved crime scene sketch dataset.Finally,we measured the performance of our Net by accuracy,ROC curve and confusion matrix,and extracted the images which were misclassified by our model.The results show that XCTNet achieved the accuracy of 98.65% on test set which is 3.8 percentage points higher than AlexNet by using one tenth parameters of AlexNet.The model achieved a high accuracy on classifying the Crime Scene Overview Sketches,MapScreenshot Location Sketches and the negative samples,and ability of classifying Self-Drawn Location Sketches needs to be enhanced.