Image Reconstruction from Local Binary Patterns

Reconstruction of an image from its LBP codes can aid understanding of the information contained within the codes by comparing the reconstructed image to the original. We are the first to show that the LBP process can be inverted and present a novel algorithm to perform the reconstruction, resulting in an approximation of the original image that is both visually appealing and completely matches the LBP codes of the original. The algorithm calculates the minimum contrast between two pixels, reconstructing some of the contrast information thought lost in the LBP process. Tests on the algorithm have been conducted on images from the Brodatz database and Berkeley Segmentation Dataset which show an image visually similar to the original with perfect texture reconstruction. The reconstructed images also remove the effects of illumination from the images, suggesting future investigation into the possibility of image brightness normalisation. Additionally, since the reconstructed image provides the same LBP codes as the original, the susceptibility to spoofing of systems using LBP feature vectors has been identified.