Scholarlink

<p>Lookup table (LUT) has shown its efficacy in low-level vision tasks due to the valuable characteristics of low computational cost and hardware independence. However, recent attempts to address the problem of single image super-resolution (SISR) with lookup tables are highly constrained by the small receptive field size. Besides, their frameworks of finite-layer lookup tables limit the extension and generalization capacities of the model. In this paper, we propose a framework of series-parallel lookup tables (SPLUT) to alleviate the above issues and achieve efficient image super-resolution. On the one hand, we cascade multiple lookup tables to enlarge the receptive field of each extracted feature vector. On the other hand, we propose a parallel network which includes two branches of cascaded lookup tables which process different components of the input low-resolution images. By doing so, the two branches collaborate with each other and compensate for the precision loss of discretizing input pixels when establishing a lookup table. Since SPLUT uniformly discretizes the input density space for the series-parallel index, the input distortion by sub-optimal index quantizers results in inefficient utilization of the non-linearities capability for each LUT. With LUTs expansion, the performance becomes increasingly constrained due to the cumulative quantization errors. To address this, we propose SPLUT with learnable index quantizers, named LISPLUT, which automatically adjusts index distributions by non-uniformly allocating the quantization thresholds according to the input density distribution with minimal storage and computational overhead. Moreover, we introduce a lightweight bit-level spatial compensation to inherit the computational efficiency of the independent parallel topology and compensate for the quantization drop by enhancing the guidance of the complementary bit-level components. Compared to previous lookup table-based methods, our framework has stronger representation abilities with more flexible architectures. Furthermore, we no longer need interpolation methods which introduce redundant computations so that our method can achieve less energy cost. Extensive experimental results on five popular benchmark datasets show that our method obtains superior SISR performance in a more efficient way. The code is available at <a href="https://github.com/zhjy2016/SPLUT">https://github.com/zhjy2016/SPLUT</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Learning Extensible Series-Parallel Lookup Tables for Efficient Image Super-Resolution

  • 期刊论文

  • 发表日期：

    2025年7月2日

查看全文
  • Jingyi Zhang,
  • Ziwei Wang,
  • Cheng Ma,
  • Jie Zhou,
  • Jiwen Lu

摘要

Lookup table (LUT) has shown its efficacy in low-level vision tasks due to the valuable characteristics of low computational cost and hardware independence. However, recent attempts to address the problem of single image super-resolution (SISR) with lookup tables are highly constrained by the small receptive field size. Besides, their frameworks of finite-layer lookup tables limit the extension and generalization capacities of the model. In this paper, we propose a framework of series-parallel lookup tables (SPLUT) to alleviate the above issues and achieve efficient image super-resolution. On the one hand, we cascade multiple lookup tables to enlarge the receptive field of each extracted feature vector. On the other hand, we propose a parallel network which includes two branches of cascaded lookup tables which process different components of the input low-resolution images. By doing so, the two branches collaborate with each other and compensate for the precision loss of discretizing input pixels when establishing a lookup table. Since SPLUT uniformly discretizes the input density space for the series-parallel index, the input distortion by sub-optimal index quantizers results in inefficient utilization of the non-linearities capability for each LUT. With LUTs expansion, the performance becomes increasingly constrained due to the cumulative quantization errors. To address this, we propose SPLUT with learnable index quantizers, named LISPLUT, which automatically adjusts index distributions by non-uniformly allocating the quantization thresholds according to the input density distribution with minimal storage and computational overhead. Moreover, we introduce a lightweight bit-level spatial compensation to inherit the computational efficiency of the independent parallel topology and compensate for the quantization drop by enhancing the guidance of the complementary bit-level components. Compared to previous lookup table-based methods, our framework has stronger representation abilities with more flexible architectures. Furthermore, we no longer need interpolation methods which introduce redundant computations so that our method can achieve less energy cost. Extensive experimental results on five popular benchmark datasets show that our method obtains superior SISR performance in a more efficient way. The code is available at https://github.com/zhjy2016/SPLUT.