X-MESMERISE expected features: A full body non-divest non-intrusive quick non-disruptive ultra-low dose x-ray scanner (FBS), usable upon consent by the final user, for internally and externally concealed commodities detection identifying the object and warning the operator: Non-divest: the final user does not need to divest outer clothes or belongings from pockets, etc. Non-intrusive: it is completely contactless and because the ATR shows the alert, it is not necessary to render and display an image to the operator Quick: the whole operation could be deployed in a very convenient layout for both the customer and the security service, taking less time than the current systems Non-disruptive: the operation takes place in a continuous moving belt maintaining a constant flow of passengers The dose per scan is under the limits according to the EU standard limits ( 1mSv per year) Automated Target Recognition System

Inspired by tumor segmentation models Train a CNN (Convolutional Neural Network) model for Semantic Segmentation.

Our model Overall Architecture: Deep Encoder-Decoder CNN-based (Convolutional Neural Network CNN) Given a feature representation obtained from the encoder network, dense pixel-wise prediction map is constructed through decoder network.

Our model Training Process: Optimization

Mesmerise Dataset Number of X-ray images (from London & Oslo scanners): 57 Raw X-ray image                Raw X-ray image      Mask Mask London                             Oslo

Mesmerise Dataset

Results on Mesmerise Dataset Sorensen-Dice Ground Truth   coeficient

4.1 Quantitative Results Setups: Set 1:            Training images: 49 Validation images: 8 Set 2: Training images: 29 Validation images: 28 Evaluation metrics: Global Accuracy (GA) – Pixels right Class Average Accuracy (CAA) – Proportion of Right Pixel per Class Intersection over Union (IoU) – Overlap between GT and Predicted Masks Dice coefficient (D) – Overlap with class 1 (narcotics positive)

4.1 Quantitative Results The importance of the number of training images Setups       GA (%) CAA(%)   IoU       D SET 1 (49     99.16  71.10  0.6212   0.3519 training images) SET 2 (29     99.16  64.41  0.5761   0.2725 training images)

4.2 Qualitative Results: Probabilistic Output Probability of finding narcotics in that pixel