Deep Learning for Firearm Detection from X-ray Images

Project Details


This proposal will develop an innovative machine learning approach based on deep learning CNN that can automatically recognise and identify firearms from a dataset of X-ray images.
The project will investigate, analyse, design, developed, test and validate CNN architectures and employ them for object recognition (firearms) from the given datasets, trough ‘end-to-end’ learning of hidden relationships, patterns and trends (that are not based on pre-conceived and biased ‘expert’ features), from the ‘raw’ image pixels on one end to class scores at the other. The assumption that the inputs are images, allows the encoding of certain properties into the CNN architecture, which makes the implementation of the forward function more efficient and reduces considerably the number of network parameters.
The investigated CNN architectures will receive as an input (image pixels intensity) and will transform it through a number of hidden layers, the output part of the CNN will be based on fully connected “output layer/s”, including a SOFTMax layer for a probabilistic output of the classification – giving probability confidence interval for the object belonging to a particular class. (initially two classes will be considered: ‘threat’ and ‘benign’, and subsequently a multiclass output that will include specific weapon parts, such as ‘spring’, ‘barrel’, and ‘trigger’ will be also investigated).

Key findings

Automated threat detection system based on the state-of-the-art machine learning approach capable of detecting and recognising ‘threats’ from X-ray images. Once optimally trained, the system will have readiness for on-line implementation, connecting directly to X-ray cameras.
Short titleSBRI_ ATD of Firearms Phase 2
Effective start/end date1/11/171/11/19


  • pattern recognition, deep learning, convolutional neural networks


Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.