Detection of sharp objects using deep neural network based object detection algorithm

Deep learning algorithms have the ability to learn complex functions and provide state-of-the-art results for computer vision problems. In recent times, these algorithms far exceeded the existing computer vision based techniques for object detection in X-ray imaging systems. So far, in literature single class of object namely gun and its parts were considered for detection using the SIXray10 database. We propose deep learning-based solution for the detection of sharp objects namely knife, scissors, wrench, pliers in the SIXray 10database. We propose two models namely model A and model B using a common object detection algorithm- YOLOv3 (You Only Look Once) with InceptionV3 and ResNet-50. YOLO is a deep neural network based object detection algorithm that performs the task in one-shot which allows real time inference in video of 15-30 fps. The model is FCN (Fully Convolutional Network) as has the capacity to perform both regression and classification by sharing weights for both the tasks. The network predicts a rectangular box called bounding box around the predicted object of interest along with the associated class. We analyze the performance of both model in terms of mAP. We achieve mean accuracy of 59.95% for model-A and 63.35% for Model-B. The most daunting part of the project is the low ratio of harmful to nonharmful items. By performing rigorous experiments we came up with the best set of possible results which uses varied pretrained neural networks for feature extraction in tandem with YOLO model for object detection. We endeavor to improve on these existing results so as these systems can be successfully deployed in airports to minimize human error and improve security in such environments.