ProTEcTME: Proposal based exTEnsible Threat detection fraMEwork

Project Description

The automatic identification and classification of single objects within an image representation of a scene, is in itself a challenging computer vision problem. Previous research has achieved good results in simple structured images. However, objects in cluttered, noisy images, where they might be occluded or obscured, are much harder to detect. Very robust detection of objects in real-world situations is still a wide and open area of active computer vision research.

Most of the currently used image recognition techniques rely on pre-trained image sets where ‘target’ objects are clearly visible and defined. This has drawbacks in that small variations in the object can prevent identification or classification; techniques such as image flipping and discrete rotations have aimed to solve this, but are only partially successful.

This project provides a robust solution to the automatic identification and classification of 'target' objects via cuttingedge advances in Convolutional Neural Networks (CNNs). Unlike more traditional methods of computer vision object recognition, the strength of the solution provided here is that the system learns to identify the object by inference by simply observing images that contain the object of interest. The system therefore learns and is analogous to a human brain, where we learn by repeated observation to identify features of an object. The advantage of this approach is that we can learn new objects by observing them, rather than programming or engineering the system to recognise the features of the object.

This system also takes advantage of modern day Graphics Processing Units (GPUs) providing incredible speed-ups in performance due to the parallel-nature of Neural Networks. Once a network is trained, these can be invoked and an image detection and classification obtained as quickly as 1/100th of a second.

Such a system provides a robust threat detection solution, whereby 'target' objects can be identified and classified in real-time; in addition to providing a flexible and adaptable environment required in an ever-changing threat landscape. The system has the ability to learn new 'target' objects by simply providing new positive images of the object to be detected, simply by providing these images from an operation pipeline, even noisy and occluded images, the more realistic and operational the better.