Integrating dilated reparameterization with lightweight multi head self-attention for iris segmentation and liveness detection

In response to the multiple challenges in iris segmentation and liveness detection tasks: First, existing research predominantly relies on classical convolutional neural networks and Transformer architectures, causing the propagation of their limitations in local-global focus dynamics and architectural complexity to downstream tasks. Second, the inherent integration of iris segmentation and liveness detection as subtasks within a unified framework necessitates a lightweight network design. Third, robust processing capabilities are required in non-cooperative environments. Fourth, limited data conditions impose constraints on model training. This paper proposes a backbone network that integrates dilated reparameterization with lightweight multi head self-attention, specifically optimized for iris segmentation and liveness detection. Dilated reparameterization employs multi-branch architectures during training, and switches to single-branch architectures during inference. By merging parallel small kernels to optimize large-kernel parameters, this method mitigates the limitation of convolutional neural networks for iris image processing that overfocus on local texture features, achieving multi-scale perception and improving computational efficiency for related tasks. The non-adjacent flattened tokens method preserves long-range associative information between iris and non-iris regions, enabling lightweight multi head self-attention to extract more discriminative iris features. Additionally, we propose a combined data augmentation method to enhance the generalization ability of the model, and demonstrate its effectiveness through multivariate classification and cross validation. The experimental results demonstrate that the proposed framework achieves significant performance advantages compared to state-of-the-art methods based on classical convolutional neural networks or Transformer architectures, while requiring only 13% of the average computational overhead of mainstream methods.