Improving Object Recognition in Crime Scenes via Local Interpretable Model-Agnostic Explanations

Object recognition is one of the fundamental and challenging problems in most visual monitoring applications and security systems. In assessing of crime scene, the detected objects from an image play a vital role in providing a narrative summary of a scene. It helps police officers in the investigation of the crime scenes and evidence collection in the criminal justice system from the viewpoints of law enforcement. In recent years, deep learning methods have emerged as a powerful strategy for the recognition and detection of objects in images. However, still, there are potential problems in the use of machine learning and Artificial Intelligence (AI) algorithms in policing. For instance, when using deep learning-based object detection approaches in crime scenes there are risks in that algorithms when making predictions. Because the complex structures of deep learning do not allow even the educated users to understand the mechanisms behind its decision-making processes. In this paper, we demonstrate how to enhance object recognition in crime scenes by utilising Local Interpretable Model-Agnostic Explanations (LIME). We use LIME to identify the object’s parts that have the highest impact on the class score and to show which critical features are employed in decision-making. To this end, the LIME heatmap is used that provides us valuable information that photography with which angles and perspectives can reduce the weak spots of deep learning-based object detection in the crime scenes. The platform has been tested with a wide variety of images of crime scenes. The results of our experiments provide support for our hypothesis.