Image analysis for blood spatter problems

Abstract

Violent crime scenes are becoming increasingly common nowadays. Usually in such cases, the obvious evidence of the crime is blood spatter. Forensic specialists try to predict the event of the crime as accurately as possible based on blood spatter evidence in the scene. Recently, programs have also been developed to predict the events in the crime scene. However, there are several shortcomings including predicting the source of origin and trajectory of the blood drop, complications from large amount of manual input and lack of research on related classification methods, such as Neural Network (NN) in this field. In this thesis, focus is given to enhance the prediction method both theoretically and practically. The proposed theoretical model is based on the Newton’s Law for linear blood spatter drop in motion, assuming the motion has drag. It produces more accurate results compared to the model using Stokes’ Law, which has been used in previous researches, if blood droplet radius is more than 2 mm, otherwise they are comparable. To perform experimental research, a number of available blood stain image data is necessary, but there is no available data. Hence, a database (DB) with 1252 blood stain images has been created through the formation of synthetic blood formula and practical bloodletting crime image scenario. Finally, the classification and automation for the reconstruction of blood droplet trajectory using two different Neural Networks (NN) modules which are Cascade Forward Neural Network (CFNN) and Function Fitting Neural Network (FFNN) is proposed. The CFNN and FFNN then tested with the developed image data-set. FFNN exhibits in average 91.1% classification accuracy for blood stain images, which is 4.5% better than CFNN and significantly better than previous researches. The proposed system may help forensic investigators to acquire crime scene evidence in an easy, faster and reliable way in near future.