Insertion and deletion (indel) error reduction using deep-learning segmentation for DNA data storage

This project implements a pipeline for predicting and reducing indel (insertion–deletion) errors in DNA-encoded image sequences using deep-learning–based segmentation models, alongside substitution-oriented noise reduction through image denoising. Three segmentation networks—UNet, DeepLabv3, and SegFormer [1–3]—are evaluated for their ability to detect indel positions from generated segmentation maps, while a DnCNN [4] model is employed to suppress substitution noise in the decoded images. Input images are grayscale samples derived from the STL-10 dataset located in "data/input/". The complete IDS (i.e., insertion, deletion, and substitution) error–reduction workflow proceeds as follows: Input grayscale image → Byte-to-binary conversion → DNA encoding (binary-to-DNA sequence) → IDS noise injection → DNA decoding with segmentation-based indel reduction (noisy image → segmentation → indel correction) → DnCNN denoising. Install Requirements pip install -r requirements.txt Running python main.py--input_image ./data/input/01.png--output_dir ./data/output/--encoding_period 16--sub_ratio 0.005--indel_ratio 0.00003--seg_model unet--device cuda Outputs [1] Encoded DNA sequence: ./data/output_image/01.dna [2] IDS-noisy DNA sequence: ./data/output_image/01_noisy.dna [3] Noisy image (before indel reduction): ./data/output_image/01_noisy.png [4] Indel-reduced image (segmentation-guided): ./data/output_image/01_indel_reduced.png [5] Denoised image (DnCNN): ./data/output_image/01_denoised.png References [1] Ronneberger, O., Fischer, P., Brox, T. U-Net. LNCS 9351, 234–241 (2015). [2] Chen, L.-C., et al. Encoder–Decoder with Atrous Separable Convolution. ECCV (2018). [3] Xie, E., et al. SegFormer. NeurIPS 34, 12077–12090 (2021). [4] Zhang, K., et al. DnCNN. IEEE TIP 26, 3142–3155 (2017). Notes The implementation uses Rinf_encoding/decoding (4-bit to dinucleotide mapping with periodic shift). Segmentation weights are loaded from ./models/*.pth; SegFormer may download base weights on first run (see requirements). STL-10 copyrights and model weight licenses are separate from this repository’s license.