Efficient Image Super-Resolution with Collapsible Linear Blocks

In this paper, we propose a simple but effective architecture for fast and accurate single image super-resolution. Unlike other compact image super-resolution methods based on hand-crafted designs, we first apply coarse-grained pruning for network acceleration, and then introduce collapsible linear blocks to recover the representative ability of the pruned network. Specifically, each collapsible linear block has a multi-branch topology during training, and can be equivalently replaced with a single convolution in the inference stage. Such decoupling of the training-time and inference-time architecture is implemented via a structural re-parameterization technique, leading to improved representation without introducing extra computation costs. Additionally, we adopt a two-stage training mechanism with progressively larger patch sizes to facilitate the optimization procedure. We evaluate the proposed method on the NTIRE 2022 Efficient Image Super-Resolution Challenge and achieve a good trade-off between latency and accuracy. Particularly, under the condition of limited inference time (<= 49.42ms) and parameter amount (<= 0.894M), our solution obtains the best fidelity results in terms of PSNR, i.e., 29.05dB and 28.75dB on the DIV2K validation and test sets, respectively.