Computationally-Efficient Neural Image Compression Using Asymmetric Autoencoders

Neural image compression (NIC) is an emerging field, with notable works achieving promising results in image quality. However, these networks introduce new challenges regarding computational cost since NIC require high computational power for compression and decompression, demanding high processing time even when specialized platforms like GPUs are used. This work proposes optimizing one of the reference models available in the literature to achieve better processing time while maintaining the compression quality. Unlike other solutions that aim at higher compression or better image quality, our proposal seeks to reduce the computational cost of NIC techniques. This is achieved by designing novel asymmetric autoencoder architectures with cost-efficient versions of the decoder component of these networks. Experimental results obtained with the Kodak dataset show that, by pruning layers or kernels at the decoder network, it is possible to achieve a 18% reduction in decompression time for GPU, with an average PSNR loss of 0.001 dB. When tested on a CPU, the decompression time reduction was 46%, with a PSNR loss of 0.007 dB. We expect that our contributions will pave the way for more efficient NIC techniques, making them more accessible for real-time applications and resource-constrained environments.