Introduction
Neural Image Compression (NIC) has achieved state-of-the-art rate-distortion performance, often outperforming traditional standards like JPEG and BPG. However, these gains come at the cost of high computational complexity, frequently exceeding 50 GFLOPs. Such demands hinder deployment on resource-constrained edge devices. While unstructured sparsity can reduce weight counts, it requires specialized hardware for speedups. This work explores structural sparsity through channel pruning, enabling static, "slimmer" networks that accelerate inference on standard hardware.
Methodology: Adaptive Binary Channel Masking (ABCM)
The authors introduce the Adaptive Binary Channel Masking (ABCM) module, a plug-in layer designed to identify and remove redundant convolution channels.
The process involves three main stages:
- Masking: A learnable vector $\alpha$ determines channel importance. During training, a gating function (approximated via Sigmoid) masks unimportant channels by setting them to zero.
- Sparsity-Driven Loss: A penalty term is added to the loss function to minimize the number of active channels, encouraging the model to compress information into a subset of the network.
- Static Pruning: Post-training, channels with zeroed masks are physically removed, resulting in a permanently reduced architecture that requires no specialized hardware.
The models were trained on the CLIC dataset and evaluated using the Kodak benchmark across three popular NIC architectures: Factorizedprior, ScaleHyperprior, and JointAutoregressive models.
Results and Analysis
Experiments reveal significant redundancy in trained NIC models. In some configurations, nearly 50% of channels can be pruned with only a minor drop in Peak Signal-to-Noise Ratio (PSNR).
Computational Efficiency: The Factorizedprior model achieved up to a 7.17x reduction in FLOPs and a 5.63x reduction in parameters. On standard hardware, this translated to a 3.40x speedup on an Intel Xeon Gold CPU and 2.33x on an NVIDIA Tesla V100 GPU.
Rate-Distortion Trade-off: These efficiency gains were achieved with a negligible PSNR drop (typically <0.5%), making the visual difference virtually imperceptible. Interestingly, models optimized for lower bitrates exhibited higher redundancy and could be pruned more aggressively than high-quality models.
Structural Insights: While the main encoder and decoder are highly prunable, the entropy-modeling "hyper-codec" proved sensitive to pruning. Pruning this component led to significant performance drops, so it was kept intact in the final optimized models.
Conclusion
ABCM provides a simple, effective "plug-and-play" solution for optimizing NIC models. By leveraging structural sparsity, researchers can significantly reduce the computational footprint of neural codecs without sacrificing visual quality. Future work will investigate joint optimization of the entire pipeline, including the hyper-codec, to further push the boundaries of efficiency in real-time applications.
Summary updated