Aluminum Casting Inspection using Deep Learning: A method based on Convolutional Neural Networks

Abstract

In the last years, many computer vision algorithms have been proposed for baggage inspection using X-ray images. In these approaches, the idea is to detect automatically threat objects. In baggage inspection, however, a single view is insufficient because there could be occluded parts or intricate projections that cannot be observed with a single view. In order to avoid a misinterpretation based on a single view, we propose the use of mono-energetic multiple X-ray views. Our approach computes a 3D reconstruction using Space Carving, a method that reconstructs a 3D object from its 2D silhouettes (that have been segmented using Geodesic Active Contours). The detection is performed by analyzing 3D features (obtained from the 3D reconstruction). Instead of dual energy, that is typically used in baggage inspection to analyze the material of the reconstructed objects, we propose to simply use mono-energy for the detection of threat objects that can be recognized by analyzing the shape, such as handguns. The approach has been successfully tested on X-ray images of travel-bags that contain handguns. In the evaluation of our method we have used sequences of x-ray images for the 3D reconstruction of objects inside travel-bags, where each sequence consists of 90 x-ray images. we obtained 0.964 in both precision and recall. We strongly believe that it is possible to design an automated aid for the human inspection task using these computer vision algorithms.