Currently, the results of iris recognition cannot be applied to judicial trials. The forensic science community has begun to focus on quantitative identification method based on the statistical rules of interpretable iris features. For this purpose, it is necessary to achieve automatic segmentation and extraction of iris texture features. A segmentation network for block-shaped iris features in near-infrared iris images is proposed, which combines residual networks, attention mechanisms, and atrous spatial pyramid pooling. First, a block-shaped iris feature annotation dataset was constructed for model training, validation, and testing. Secondly, improvements were made to the UNet framework as follows: the convolutional modules were replaced with residual modules to promote gradient propagation and enhance feature retention and transmission capabilities; attention gate modules were added to the skip connections to improve the model′s perception of block-shaped features; and atrous spatial pyramid pooling modules were employed in the middle and end parts of the model to expand the receptive field and perform multi-scale feature extraction and fusion. Finally, a hybrid loss function combining cross-entropy and Dice coefficient was proposed, and preprocessing techniques such as normalization and histogram equalization were used to highlight block-shaped iris features. Experimental results show that the RAA-UNet outperforms other comparison models on the test set, with F1 score, mIoU, and Dice coefficient reaching 82.64%, 84.21%, and 81.66%, respectively, representing improvements of 4.42%, 3.37%, and 3.87% over UNet. The loss function experiments determined the optimal weight factor, and histogram equalization significantly improved segmentation performance. Ablation experiments verified the effectiveness of the model improvements. The proposed RAA-UNet semantic segmentation model can accurately segment block-shaped iris features, providing technical support for iris identification research.