Artificial Intelligence (AI) X-Ray Threat Detection Technology for Counter IED Operations

Artificial Intelligence (AI) threat detection software is becoming more and more viable in the cabinet x-ray market. The ability of AI software to automatically detect threat objects is becoming very realistic as the software has improved. Most of the systems currently on the market have focused on detection of the gun and knife type threats but the IED is still a "work in progress". This is because of the overall complexity of the IED threat and all of the possible variations. It will be very challenging to develop an AI technology that can provide accurate automatic detection for the IED type threat. While it is not impossible we are just not there yet with an off the shelf AI auto detection system for IED's. With Portable x-rays used by counter IED Bomb Disposal professionals they are extremely far behind when it comes to the type of auto detection capabilities we currently have with cabinet x-rays. AI detection is becoming a reality with cabinet systems but can we implement a similar capability in the portable systems?

All of these AI software developments are focused on the cabinet x-ray market and only very limited efforts have been made to implement this type of technology for Bomb Disposal x-ray systems. One of the problems with doing any type of automatic detection with the portable systems is that they are using x-ray technology that is just not able to work outside of grey scale imaging (single energy). The pulsed based systems that most bomb teams use cannot even come close to the capabilities that a cabinet x-ray system can provide. Dual energy imaging is the standard in the cabinet x-ray industry but with pulsed based portable systems all efforts to develop accurate materials discrimination imaging capabilities won't pass established standards. Until Bomb Disposal moves away from pulsed based systems and embraces constant potential x-ray generators that the rest of the security x-ray market is using they will be stuck with black and white single energy imaging capabilities.

In the above image you see what is typical with a pulsed x-ray system trying to do a materials discrimination image compared to a system using a constant potential x-ray generator. If you place a test object with an organic, inorganic, and metal disc in different locations on the aSi panel you will get different readings with a pulsed generator. This makes a pulsed portable systems ability to provide "accurate" materials imaging very suspect and honestly dangerous. However most Bomb Disposal have no idea how materials imaging works so they are none the wiser about what they are looking at is not accurate. This inability really limits the pulsed based systems potential for becoming a better tool and developing any new capabilities that we see in the cabinet x-ray systems. Bomb Disposal seems to be stuck in the medical x-ray world and honestly have no real understanding of what other technologies are available. This inability to do intelligent market research is one of the biggest reason they are still using single energy technology.

The next example is very indicative of how bad a pulsed based x-ray using a aSi detector is at materials discrimination imaging. Pulsed x-ray generators are a very unstable tool for trying to do accurate materials imaging and as you move the generator up and down or backwards or forwards the measurements change drastically. Bomb technicians do not understand this or typically what the colors mean and take the sales pitch from these vendors hook line and sinker. If you do not believe me reach out to me and I will send you a test kit for free and your pulsed system will fail (Trust me). When I see Bomb units who have purchased a dual energy module for a pulsed generator I know they have been ripped off.

There is however still an application for AI autodetection with grey scale images. The AI technology being used today is only looking for pixel data to automatically detect specific threat items. Pulsed based systems will never be able to do something like automatic detection for explosives (density and Z eff measurements) like cabinet system can but they can implement AI detection for IED components using shape recognition. To accomplish this is not going to be a simple task and requires a exceedingly large amount of effort to accomplish vs gun and knife AI detection. Lets take a look at some of the challenges involved with AI automatic detection for the IED type threat using portable x-ray technology.

Problem 1: Massive Number of Potential Components

As a Bomb Technician myself I know there are a "OMG" amount of different ways to make a IED. This fact is why the development of IED component based IED AI detection is not even close to the level of gun and knife AI detection. Guns and Knives are easy and you do not need to be a IED "Expert" to figure them out. When presented with IED's, and trying to develop a AI detection capability it can appear to be almost impossible. Even if you did develop a list of all of the different parts and pieces that can be used each item would require many subcategories. This is because you can have many different variations of each IED component. The below image is a perfect example looking at just switches and how many different ways they are constructed. Also note this is a very small sample of all of the different types of switches that can be used in a IED.

Problem 2: Varying KV/Pulses (AKA Scan Time) effects the image and pixel values

With portable x-ray systems the the Bomb Technician can adjust the pulses and this will effect how a item will look in a x-ray image. At higher pulses and KV you can see lower density items wash out and disappear. The will create a completely different pixel range for the software to try and identify for each IED component. Cabinet x-ray systems do not have this problem because they always operate at a set KV and scan time. The cabinet x-ray detectors are also always getting the same amount of x-ray energy where a aSi panel energy spread is all over the place. These two factors alone makes developing a AI detection for portable system very challenging and costly. The below is a example of the wash out effect on IED switches and how the varying KV or scan time can effect the pixels. At the low KV the lower density plastics are visible so if we taught the AI these pixel values for switches and tested it with a image at a higher KV it probably would not autodetect. The metal of the switches does not wash out so with portable x-ray using AI detection you would have to take this into consideration. You would have better luck building a AI detection pixel set around the metal of the switch vs the plastic components.

You need a ton of reference angles of each item

For any AI detection software to work it has to have a extensive set of reference images of that item at all of the different angles it can be orientated. Checkpoint screeners use Fictional Threat Images (FTI's) to learn IED's and at a minimum you need 8 or more different orientations for a human to learn a specific threat in a x-ray image. With AI you are going to need more orientations and you could guess based on the item it could have a different pixel range in 30 or more orientations. The below is a example of a IED at many different angles and you can see that it looks much different at all of these different angles. Now add to this what we just discussed about varying kV and scan times and the number of references you need for a AI to learn a specific item increases far beyond what a cabinet system would require.

Clutter adds to the complexity of AI detection

If you just input 100 different references images of a IED circuit component into a AI software it will work great. But when you add bag clutter to that item and you start getting overlap the process becomes much more complex for detection. Guns and Knives are typically at a higher density range vs IED components so if they are using pixel color as a tool to streamline detection they can lower false alarms and even improve detection in clutter. Because pulsed based portable system cannot do accurate materials imaging you would never be able to implement this type of materials discrimination data . Pixel color in a materials image is basically how they show you the Z eff (average effective atomic number) of that pixel. A light orange pixel is a lower Z eff vs a darker orange pixel. This is a very easy way to explain how they do explosive detection and threat mass by looking for a specific color range and a cm2 pixel size. X-rays cannot tell you weight they determine threat mass based on pixel (or voxel) range. Pulsed based portable x-rays are just not up to the task for this level detection so clutter will be a issue for implementing AI in a portable x-ray. The other problem is Bomb technicians avoid clutter like the plague and when they train you rarely seem them use any clutter in their training IED's. This would lead to development of any AI by a Bomb Tech organization to build a AI image base solely based on no clutter. When tested in the real world it would be very limited when clutter was part of the package.

Part of the clutter problem is how IED's are actually built by terrorist. They build them in a holistic configuration meaning all of the IED circuit component are combined and attached to the explosive main charge. Bomb technicians have a very bad habit of designing their training aids not holistically but all laid out with each component separated. Detection with a AI technology of a IED and it components would be easy if it is all laid out. However when that same IED circuit is in a holistic configuration the AI software would be hard pressed to detect the components if it was taught a laid out design.

In the above image you see the exact same IED circuit and the only difference is one is all laid out and the other is holistic. To a AI system these are two entirely different learning processes. If you teach the AI system the laid out design it would be hard pressed detect the same circuit in the holistic configuration.

Can we Simplify The process Somehow?

In my "Opinion" I think that AI detection with portable systems can be accomplished but it is going to take the Bomb Technician community getting much smarter about what x-ray technology is available outside of the pulsed based medical detector panel equipment they are using now. Until they move away from pulsed generator and medical detector panels there is not much advancement they can accomplish. However there are newer portable x-ray systems on the market that have embraced the advancements in the cabinet x-ray world and have moved away from single energy imaging. These are the constant potential x-ray systems that use line scan detector technology. They are leaps and bounds more capable vs the pulsed based aSi medical panel system and have the ability to do automatic detection and accurate materials discriminations imaging.

The best place to start with AI detection for IED components would to be to look at IED's and pick a very common component that is used in almost all IED's. It would need to be something that would not have many different variations and easy to develop a reference library. The most common item would be a power source or battery.

Batteries are common in IED and most are all electrically initiated. They also have a standard size and shape which would make them an ideal candidate for developing a AI detection reference library. They also have a stable density and Zeff ranges that could also be referenced for pixel color algorithms (grey scale or materials discriminations). Out of all of the different IED circuit components in my opinion the power source would be the best place to start not only for portable x-ray system but also the cabinet x-rays.

The next very common thing that most all IED's have is a detonator. The detonator is typically going to be inserted into the explosive main charge. If you can find the detonator you have a really good idea where the exploisves are in the x-ray image. Detonators come is varying lengths and configurations but they all have one thing in common that is a very unique x-ray characteristic. They almost all have a lead based primary charge (Lead azide/Lead Styphnate) that shows up as a black mass inside of the detonator shell. This is unique because it blocks x-ray and is in a very small pixel range. A simple algorithm that looks for a small and specific pixel range (MM2) and has 100% absorption could potentially be effective at detecting these. I have actually tried this on a cabinet x-ray by adjusting the high density detection to a very small cm2 range and setting the absorption to 100%. It worked 100% of the time that I tested it and it even worked with clutter in the package.

The reality is that portable counter IED x-ray systems are not even in the small ball park when it comes to the capabilities of a cabinet x-ray system. The only people that are to blame for this are the Bomb technicians themselves for not taking the time to see what advancements have been made outside of the medical x-ray technology they seem to be stuck using. Until Bomb Disposal gets much smarter about x-ray technology they will be stuck in a black and white world but all the while thinking they are the smartest guy in the room. There is a ton of room for improvement but like the old say goes "You can lead a horse to water, but you cannot make it drink".

If you are interested in seeing cutting edge advanced portable x-ray technology click on the below link. 3DX-RAY has developed cutting edge portable x-ray systems that is not medical x-ray technology. They have embraced all of the benefits that you can find in the cabinet x-ray world and incorporated them into a portable x-ray system that is the most advanced in the world. Schedule a demo today and see for yourself what you are missing.