无块效应的基于块的学习图像压缩
Block-based Learned Image Compression without Blocking Artifacts
给出一个解析框架，通过卷积/转置卷积层推导重叠传播，精确算出无块效应重建所需的最小重叠量，可直接套用到任意 CNN 图像处理网络（超分、去噪、复原）做分块推理，在不重训练的前提下把峰值显存和计算量降到约 13% 和 2%。
原文依据 an analytic framework that models overlap propagation through convolution and transposed convolution layers to precisely determine the minimal overlaps required for blocking-free reconstruction
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论学习图像压缩（LIC）优于传统编解码器，但在处理高分辨率图像时存在峰值内存占用过高的问题。因此，基于块的 LIC 被研究用于降低峰值内存与峰值计算开销，但它常常引入块效应，导致视觉质量下降。为缓解这一问题，JPEG-AI 标准引入了一种基于图像块的方案，其中重叠块被独立编码，重叠尺寸由经验确定。然而，通过实验搜索最优重叠既耗时，又无法保证无块效应的重建。本文提出一个解析框架，将重叠传播建模为卷积与转置卷积层的作用，从而精确确定无块效应重建所需的最小重叠量。基于计算得到的最小重叠，我们给出了一套适用于大多数基于 CNN 的 LIC 模型的基于块的实现方法。在将 4K 图像划分为多种块尺寸（256×256、512×512）的条件下，将该方法应用于四个基于 CNN 的 LIC 模型，我们的方法取得了与全图编码完全相同的率失真性能，同时将平均峰值内存占用降至 13.94%（编码器）和 13.33%（解码器），平均峰值计算开销分别降至 2.6% 和 1.24%。值得注意的是，所提出的基于块的框架无需对原始模型进行任何重新训练。此外，它还可应用于大多数基于 CNN 的图像处理神经网络，且不造成性能下降。
英文原文摘要 1524 字符
Learned image compression (LIC) outperforms traditional codecs but suffers from excessive peak memory usage when handling high-resolution images. Consequently, block-based LIC has been studied to reduce peak memory and peak computational cost, but it often introduces blocking artifacts that degrade visual quality. To mitigate this, the JPEG-AI standard introduced a patch-based scheme in which overlapped blocks are coded independently using empirically determined overlap sizes. However, experimentally searching for optimal overlaps is time-consuming and does not guarantee blocking-free reconstruction. In this paper, we propose an analytic framework that models overlap propagation through convolution and transposed convolution layers to precisely determine the minimal overlaps required for blocking-free reconstruction. Based on the calculated minimum overlaps, we provide a block-based implementation methodology applicable to most CNN-based LIC models. Applied to four CNN-based LIC models on 4K images partitioned into various block sizes (256 x 256, 512 x 512), our method achieves rate-distortion performance identical to full-image coding while reducing average peak memory usage to 13.94% (encoder) and 13.33% (decoder), and average peak computational cost to 2.6% and 1.24%, respectively. Notably, the proposed block-based framework does not require any retraining of the original model. Furthermore, it can also be applied to most CNN-based image-processing neural networks without performance degradation.