Iris Recognition Network Based on Full Coordinate Attention

Although deep learning methods have shown better performance than traditional algorithms in iris recognition, exploring new CNN network architectures is still valuable. In this paper, we propose a new iris recognition network architecture named FCA-Net, which leverages the full coordinate attention mechanism. Specifically, the proposed architecture introduces an attention mechanism into the ResNet Block and uses global average pooling instead of fully connected layers at the end of the network to reduce the number of parameters and prevent overfitting. Our full coordinate attention mechanism comprises two essential blocks: CH block and W block. The CH block embeds spatial horizontal coordinate information into the channel attention, capturing long-term dependency information along the spatial horizontal coordinate while preserving precise spatial vertical coordinate information, enabling the network to accurately locate the interested parts of the iris image. The W block encodes spatial vertical coordinate and generates attention through 1D convolution with adaptive kernels, without dimensionality reduction. By complementarily applying our attention mechanism to the input feature map, we enhance the representation of the regions of interest effectively. We evaluate our model on the CASIA-Iris-Thousand and JLU-7.0 datasets, and the experimental results show our method performs favorably against its counterparts.