DNA-FRSC: A robust system for DNA-based image storage via deep visual feature representations and error-resilient sequence coding

DNA data storage is emerging as a promising technology for the long-term preservation of large-scale data, and images are among the most common and data-intensive modalities. However, existing DNA-based image storage methods remain limited in terms of storage density, reconstruction quality, and the joint satisfaction of multiple biological constraints. To address these issues, we propose DNA-FRSC, a robust DNA-based image storage method integrating deep visual feature representations with error-resilient sequence coding. An encoder from a Swin Transformer-based deep joint source-channel coding model first extracts compact latent feature representations, which are then linearly quantized and mapped to DNA sequences to improve the storage density. A constrained codebook generator produces DNA sequences with both biological feasibility and error tolerance. During decoding, a two-stage error-correction module combines a top- k dynamic programming list decoding strategy with an equal-distance candidate refinement scheme to correct insertion, deletion, and substitution errors, and a training-free feature imputation method compensates for missing feature blocks caused by sequence losses. The experimental results confirm that the generated sequences satisfy biological constraints. Extensive tests on the Kodak and General-100 datasets at a 5 % base error rate show that DNA-FRSC outperforms existing methods, reducing the MSE by more than 82.7 %, improving the PSNR by at least 8.23 dB, and increasing the SSIM and MS-SSIM by more than 83.7 % and 50.4 %, respectively. These results indicate that DNA-FRSC provides a low-cost, highly robust, and biologically feasible solution for large-scale DNA-based image storage.