Pruning the Seg-Edge Bilateral Constraint Fully Convolutional Network for Iris Segmentation

Hui Zhang, Junxing Hu, Jing Liu, Zhaofeng He, Xingguang Li, Lihu Xiao

Abstract: Iris semantic segmentation in less-constrained scenarios is the basis of new generation of iris recognition technology. In this paper, we reexamined our approach iris segmentation framework, named Seg-Edge bilateral constraint network (SEN), which contains an edge map generating network which passes detailed edge information from low level convolutional layers to iris semantic segmentation analysis layers and segmentation-edge bilateral constraint structure for focusing on interesting objects. To reduce the number of network parameters, we propose pruning filters and corresponding feature maps that are identified as useless by $$l_1$$ -norm and $$l_2$$ -norm, which results in a lightweight iris segmentation network while keeping the performance almost intact or even better. A novel $$l_1$$ -norm or [ $$l_1$$ -norm, $$l_2$$ -norm] clustering based pruning method is proposed to improve pruning effect and avoid the time consuming manual design. Experimental results suggest that the proposed SEN structure outperforms the state-of-the-art iris segmentation methods, and the clustering based pruning methods outperform manual design in both compression ratio and accuracy.