IIRDoc-Net: efficient ID segmentation with learnable IIR filters for mobile and embedded platforms

Identity document localization on mobile and embedded devices is challenging due to strict constraints on memory and computational resources. We address this problem by proposing IIRDoc-Net- a lightweight neural network architecture that integrates a block of learnable Infinite Impulse Response (IIR) filters. Unlike standard convolutional layers, IIR filters efficiently expand the receptive field by using recursive features propagation, allowing the model to capture long-range dependencies without increasing computational costs. The proposed model achieves competitive segmentation quality while significantly reducing the number of operations. In particular, it contains only 36K parameters and requires 0.3 GFLOPs, which is 32% fewer operations compared to state-of-the-art methods. Experiments on MIDV-500 and MIDV-2019 datasets show that our model achieves MIoU scores of 0.953 and 0.903, respectively. While the result on MIDV-500 is on par with existing state-of-theart approaches, the model achieves the best performance on the more challenging MIDV-2019 dataset. These results make the proposed model a strong candidate for real-time document analysis tasks on resource-constrained devices.