Evaluating one stage detector architecture of convolutional neural network for threat object detection using x-ray baggage security imaging

Abstract

Neural networks can map complex functions between input and a target, and they have produced state-of-the-art results in the field of computer vision. These neural network based models have superseded the conventional computer vision algorithms for X-ray imaging. In this paper, we propose a deep neural network based solution for a subset of the X-ray imaging problem of detecting sharp items in a baggage X-ray. Existing reports were region based CNN architecture for an object detection in X-ray imaging systems. We propose Deep learning method as a Single Shot Detector (SSD) and RetinaNet, which are a one-shot technique for object detection and are able to do inference in real time 15-30 frame per seconds (fps) videos. These techniques are Fully Convolutional Network (FCN) and have the capability to do both classification and regression with the same shared weights. These networks return a bounding box around the object of interest along with the class of that particular object. This technique has been used in training single stage detectors for four objects of interest - knife, scissors, wrench and pliers. We have achieved good detection accuracy with mean average precision of a 60.5% for SSD and of 60.9% for RetinaNet using the SIX-ray10 database, which contains harmful items and non-harmful items. The ratio of number of harmful to non-harmful items is very low,making the problem a daunting one. Through various experimentations we have come up with the best possible results using various pre- trained networks as the feature extractor in tandem with these object detection algorithms. With further improvements on the achieved results, it would be possible to deploy this technique in airports to minimize human error and improve security in such environments.