Object detection in full-height human X-ray images

1. A method for detecting and recognizing hidden objects on a human body, the method comprising:

acquiring an incoming X-ray image of the human body passing through a penetrating X-ray scanner;

generating additional images based on the incoming image by performing logarithmic or saliency transformations or contrasting of the incoming X-ray image;

obtaining maps for all objects and known object classes, wherein the maps show which pixels correspond to objects and which pixels correspond to background, by passing the incoming X-ray image and the additional images through a neural network with a deep Segnet-U-Net architecture that is optimized for overlapping object detection with long skip connections before each downsampling layer of the neural network;

using the maps, identifying unknown objects in the incoming X-ray image by recognizing all objects and objects of known classes, and excluding previously classified objects from the known classes from an entire set of segmented non-anatomic areas;

segmenting the incoming X-ray image into multiple parts; and

identifying parts containing objects belonging to both the known and to unknown classes.

2. The method of claim 1 , wherein the neural network is trained by:

using a dataset of images of people passing through the penetrating X-ray scanner and/or a different penetrating X-ray scanner;

generating additional images from the dataset of images by performing the logarithmic or saliency transformations of the dataset of images;

adding logarithmic brightness values of images of the dataset of images containing individual objects to logarithmic brightness values of original human images in order to enlarge the dataset of images with forbidden objects;

using parts of the original human images to train the neural network to improve recognition of objects;

using the neural network with the deep Segnet-U-Net architecture to identify known and unknown objects and the classes to which they belong; and

training the neural network using the dataset of images and maps for all objects and known classes in order to detect and classify previously unidentified classes of objects.

3. The method of claim 2 , further comprising using only those parts of images of human bodies that contain objects and excluding empty areas for neural network training in order to enhance the detection of objects on human bodies.

4. The method of claim 1 , further comprising using inpainting to conceal objects belonging to the known classes when displaying to an operator.

5. The method of claim 1 , further comprising separating the additional images into anatomic and non-anatomic areas so that only human body contours and foreign objects are displayed, while concealing the human body itself.

6. The method of claim 5 , wherein the separating uses segmentation of the additional images into separate anatomic areas so as to identify human body parts containing objects belonging to a class of prohibited objects.