A study of some fast inpainting methods with application to iris reconstruction

The human iris is an important biometrical element that can be successfully used for person identification. Because the image acquisition process does not provide complete information of the iris, this paper studies the problem of partial iris recognition. For texture characterization of the iris we used Local Binary Patterns, Dual Tree Complex Wavelet Transform and Daugman’s IrisCode. The evaluation of the results was made using the Equal Error Rate (EER) and AUC (Area Under the ROC Curve). The missing parts of the iris images were filled-in using patch-based inpainting. We then studied the effect and results on the iris recognition process of these inpainting methods. By using simple texture features and optimizing the inpainting method we achieved faster computations with very good recognition results. The methods were tested on UPOL, a well-known iris dataset. Our conclusion is that the inpainting process yields better results than those obtained for occluded irises.