Spatial Sequence Encoding of EfficientNetB2 Feature Maps for Altered Fingerprint Detection

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Altered fingerprints are used to escape identification, so an automated system should flag them before atching. Pipelines built on handcrafted descriptors detect most altered prints but reject manygenuine ones, ndconvolutional classifiers that pool their feature maps into a single vector discard where in the print the ridge damage occurs. We propose EB2-BiLSTM, a hybrid model that reads the 9×9feature map of a fine-tuned EfficientNetB2asasequenceof81spatialpositions and encodes it with two stacked bidirectional long short-term memory (BiLSTM) layers before classification. The model is trained in two stages and selected on a validation set. On a held out test set of real prints and prints with obliteration, central rotation and z-cut alterations at three difficulty levels, EB2-BiLSTM reaches 98.58% accuracy, 98.50% specificity, a Matthews correlation coefficient of 0.9627 and an area under the receiver operating characteristic curve of 0.9983. Under an identical data split and training protocol, the stacked BiLSTM encoder raises accuracy by 0.91 points over global pooling of the same backbone and halves the number of rejected genuine prints, while attention pooling of the same positions does not give this gain. The model also exceeds a ConvNeXt-Tiny transformer hybrid that has three times as many parameters. These results show that ordered spatial context in deep features is a useful signal for screening altered fingerprints.

This publication has 23 references indexed in Scilit: