Security X-ray baggage scanners provide images based on the different levels of radiation absorption by different materials. Images captured by such scanners are inspected by a human operator, which can slow down the verification process. To speed up inspection time, computer vision and machine learning methods are increasingly being used. While object recognition has been the subject of a huge number of articles, the problem of material recognition in X-ray images still requires some work to achieve equivalent accuracy. This paper focuses on the problem of discrimination of materials into several classes, such as organic substances or metals, in images obtained from dual-energy X-ray security scanners. We propose a new multi-scale convolutional neural network (CNN) for predicting the material class, in which five different sizes of patches are implemented parallelly to balance the trade-off between the increase in the receptive field and the loss of detail. We analyze some regularization methods and activation functions and their impact on the effectiveness of our architecture. The results were compared with other popular CNN architectures and demonstrate the superiority of our solution.

Dual-energy X-ray imaging has a vast range of application in security. Luggage inspection is an essential process for an airplane or court house security as well as securing mass events. An image of a content of some package may help to figure out if there is any dangerous object inside and avoid possibly threatening situation. As the raw X-ray images are not always easy to analyze and interpret, some image processing methods like an object detection, a frequency resolution increase or a pseudocoloring are being used. Since color can be a powerful tool to improve the usefulness of an information display, we propose pseudocoloring improvement by modifying material-based approach with edge detection to fill and sharpen color layers over the image making it easier to read and analyze. We demonstrate the effectiveness of the methods using real data, acquired from a professional dual-energy X-ray scanner.

To compare noise levels in subtraction images produced by four methods of geometric alignment. 50 pairs of intraoral radiographs (Digora Optime; Soredex, Tuusula, Finland) were used in this study. Two programs were used to correct geometric differences: ToothVis 1.4 (TV) and DentalStudio 2.0 (DS). Three reference points were manually positioned in both programs (methods 1 and 2); four (method 3) and ten (method 4) reference points were positioned within TV in each of the 50 pairs of images. The standard deviation (SD) of the histogram defining the distribution of grey shades in the subtraction image was used as the statistical parameter for evaluation of homogeneity, i.e. the noise in the subtracted images. The mean and median shade of grey values were lower for images after geometric correction in TV (126.6\126.8, 126.9\126.8 and 126.1\126.7, for three- four- and ten-point alignment, respectively) than those performed with the positioning module of DS (128.7\127.5) (P<0.05). For the SD, the mean values were significantly lower with TV (4.6, 4.0 and 3.3 for three-, four- and ten-point alignment, respectively) than with DS (6.8). The range of SD values was the largest for four-point alignment with TV (0.7-15.4), smaller for three-point alignment with DS (1.5-15.4) and three-point alignment with TV (0.5-13), and the smallest with ten-point alignment in TV (0.5-8.7). The SD of the grey-shade histogram showed that subtraction images produced with ToothVis 1.4 software were statistically less noisy than images produced with Dental Studio 2.0 software. There is a relationship between the number of reference points chosen and the noise in the subtraction images.