부산대학교 도서관

The quality problem of finger vein image largely affects the finger vein recognition performance. In order to precisely select the high-quality images, this letter proposes a reverse attention-based multi-feature interaction network for finger vein image quality evaluation. In the proposed network, the reverse attention module is used in two identical branches to extract the quality features from grayscale finger vein image and its binary vein pattern image. Additionally, the quality features from each branch is enhanced by interacting with the quality features from another branch in the proposed interaction module. Finally, the compact bilinear pooling performs the product fusion of two kinds of enhanced quality features and reduces the dimensionality of the fused features. We manually and algorithmically label the image quality of the open finger vein database from Shandong University. The classification accuracies of the proposed network are 91.67% and 86.67% on the quality-labeled images, which is superior to the performance of the state-of-the-art methods.