Exploring Asymmetric Autoencoder Architectures for Computationally-Efficient Neural Image Compression

Neural image compression (NIC) employs neural network models called autoencoders to compress images. NIC offers comparable or better rate-distortion performance and visual quality than traditional methods but requires significant processing power for compression and decompression. This work optimizes a reference NIC model. We reduce computational costs by designing asymmetric autoencoders with efficient decoders. Results on the Kodak dataset show that pruning layers in the decoder lead to an 18% reduction in decompression time on GPUs and 46% on CPUs, while requiring between 3.4% and 5% more bitrate to for the same image quality. We hope this work contributes to making NIC more efficient and accessible for real-time applications and resource-constrained environments.