Current segmentation methods do not account for the segmentation accuracy and efficiency of visible and near-infrared heterogeneous noisy iris images; thus, in this study, we propose a codec network based on the fusion of attention mechanism and dense multiscale features. First, an improved residual bottleneck element based on deep separable convolution was introduced to reduce the number of parameters and computation while preventing information loss and gradient confusion. Second, the dense void space pyramid module's void rate combination was improved and placed behind the encoder to improve multiscale feature fusion. Finally, to improve the resolution of noise targets and iris pixels, an efficient parallel space-channel attention module was designed and integrated into each down sampling layer and decoder. The experiments conducted on three open iris data sets show that both the average F1-score and mean intersection over union (mIoU) of the proposed network are superior to the existing algorithms. Compared with the benchmark network, the occupied space, number of parameters, and amount of computation are reduced by 41%, 41.77%, and 65.35%, respectively. It can effectively improve the segmentation performance of the network for multispectral noise iris and is easier to be deployed on mobile devices, which can more efficiently and accurately distinguish noise and iris targets.