Despite recent advances in computer vision humans still perform recognition of a novel scene in a single glance better than the best of the available systems. Consequently in order to achieve a similar ability in artificial intelligent systems, it is necessary to further study the low-level mechanisms in image processing for solving computer vision problems. The purpose of this study is to find an effective approach to classify images into threatening and non-threatening categories. Some of the existing algorithms for scene classification are examined and are studied in order to identify which is the best for the threatening context. We define a threat as a cause of harm or danger from a person or some phenomenon. We have constructed an image database containing hundreds of images labeled and divided into threatening and non-threatening categories. The results of classification shows that using some of the current state of art features and scene descriptors, the accuracy of classification is up to 80%.