Ozone-depleting substances and hydrofluorocarbons contraband detection within X-ray cargo images using AI
The illegal trade of ozone-depleting substances (ODS) and hydrofluorocarbons (HFCs) is a serious environmental concern. Despite regulations, recent seizures across Europe highlight a growing problem[1]. The successful interception of these products underscores the importance of vigilant monitoring and coordination among authorities in combating the illegal trade of substances that contribute to climate change.
One possible avenue of identification is via X-ray scanning of cargo containers and trucks. This identification still proves itself challenging, however, due to the high volume of goods to be processed and the requirement of human verification, it is thus important to have effective analysis tools to automate this threat detection process.
With the aid of modern Artificial Intelligence (AI) and machine learning techniques, we gain the ability to efficiently and systematically detect such threats. Specifically, we use a Convolutional Neural Network (CNN) model structure to achieve our goal.
The Analysis Technique
Starting from an input image, each layer performs a mathematical operation on the previous one, identifying more and more characteristics (shapes, colors, patterns, etc.), until the last layers gathers all the information and uses it to identify all of the desired objects (if present) from the image. To properly identify objects, a network needs to be trained beforehand on what to look for.
A Neural Network (NN) is a machine learning technique that mimics the neuron connections of a brain and allows, using a specific training process, to create a “model” able to recognize desired characteristics from a specific input.
A “Convolutional” NN (the one we use in this context) is specifically optimized to efficiently process images (as well as other visual media), by iteratively grouping sections of the input and extracting features (like shape, color and other patterns) to then identify objects based on its training. (see fig.1)
As we want to distinguish objects inside cargo containers and trucks (fig.2), the training process for the CNN makes use of a high volume of known images with and without contraband items (ODS and HFCs canisters, in our specific use case)
Once finished training, if performed properly, the model generated becomes able to recognize these items with great precision and in real-time.
End-User Experience
The final tool will be available for use on the Perivallon platform website, where users will be able to upload an X-ray cargo image of their choice that will be analyzed in real time by the CNN model and the results (should a contraband item be found) will be drawn out on the image and written in a JSON file with the coordinates and type of items found, both easily downloadable for inclusion in reports.
[1] https://eia-international.org/climate/illegal-trade-in-refrigerants/
[2] Wu, Jiatu. (2018). Complexity and accuracy analysis of common artificial neural networks on pedestrian detection. MATEC Web of Conferences. 232. 01003. 10.1051/matecconf/201823201003