Presentation Information
[N-1-39]Bit-Flip Robust Compressive Autoencoder for Image Encoding
◎Koki Watanabe1, Ryo Yamatomi1, Hiroshi Ninomiya1 (1. Shonan Institute of Technology)
Keywords:
Compressive Autoencoder,Bit-Flip,Autoencoder,Image Encoding,Decoding
This study proposes a Bit-Flip Robust Compressive Autoencoder (BFR-CAE) for image coding in noisy communication environments. Conventional Compressive Autoencoders (CAEs) encode an image into a quantized latent representation and reconstruct it using a decoder. However, bit flips caused by channel noise can corrupt the transmitted latent representation and degrade reconstruction quality.
To improve robustness, the proposed method introduces bit flips into the quantized latent representation during training. The model is optimized using both the normal reconstructed image and the image reconstructed from the bit-flipped representation by minimizing their mean squared errors from the input image.
Experiments were conducted using DIV2K for training and Kodak for evaluation. The bpp was set to 2, and the bit flip probability was varied from 0% to 10%. The results show that BFR-CAE suppresses reconstruction changes more effectively than a standard CAE, especially under high bit flip probabilities. This indicates that the proposed method improves reconstruction stability against channel noise.
To improve robustness, the proposed method introduces bit flips into the quantized latent representation during training. The model is optimized using both the normal reconstructed image and the image reconstructed from the bit-flipped representation by minimizing their mean squared errors from the input image.
Experiments were conducted using DIV2K for training and Kodak for evaluation. The bpp was set to 2, and the bit flip probability was varied from 0% to 10%. The results show that BFR-CAE suppresses reconstruction changes more effectively than a standard CAE, especially under high bit flip probabilities. This indicates that the proposed method improves reconstruction stability against channel noise.