Palmprint Recognition with Multiple Correlation Filters Using Edge Detection for Class-Specific Segmentation

We present a new series of results that show the competitive performance of advanced correlation filter classifiers for palmprint recognition. We design multiple correlation filters in subregions of the palmprint for each class. We propose a segmentation stage that selects palmprint subregions to train the filters in a class-by-class basis using different edge-detection operators. This effectively guides the filter training process to rely on regions that have a stronger line content, increasing between-class separation of the palm-prints. We evaluate the proposed algorithm in a large palmprint database of 385 classes. Our preliminary results show that most classes can be perfectly separated and the average equal error rates are as low as 0.0003% for regions of interest of size 64 times 64 pixels.