PIXELCODE: A VQ-VAE Based Image-to-Alphanumeric Reversible Encoding System

This paper presents a system, PIXELCODE, where we can preserve data by converting data into human readable alphanumeric strings. This paper demonstrates a novel reversible way to convert image-to-alphanumeric strings encoding system that support Vector-Quantized Variational Auto encoders (VQ-VAE) to compress 256×256 dimension pixel-art images into compact strings under 500-character base-62 alphanumeric strings, which is suitable for physical archival storage method and also a device-independent data preservation technique. This method achieves an extreme compression ratio of 594.96:1 while maintaining noticeable and acceptable reconstruction of quality sufficient data for content identification and archival retrieval. PIXELCODE comprises a VQ-VAE encoder-decoder architecture with discrete codebook quantization of 2,048 embedding’s, a deterministic base-62 character mapping layer implementing gzip compression and SHA-1 checksums for error detection, and a fully reversible decoding pipeline enabling bit-perfect recovery. It is locally trained on 1,014 pixel-art images (all 256×256 dimension PNG format images) which demonstrates stable convergence with 92.2% loss reduction over 120 epochs, which achieves a final mean squared error of 0.1576. This technique enables complete image to be encoded into manually documentable strings, which is feasible for paper-based storage, QR code embedding with redundant error correction, or metal engraving on durable physical surfaces suitable for multi-century preservation. Experiment validation confirms the feasibility of using neural compression for human-transferable, alphanumeric-encoded visual archival.