用卷积网络快速恢复图像DCT系数的符号，计算开销极低。
Fast Sign Retrieval via Sub-band Convolution: An Elementary Extension of Binary Classification
- 将DCT系数按频带分块成3D表示，用二分类模型预测符号。
- 在多个数据集上实现接近100%准确率，推理速度远超传统方法。
- 适合需要高压缩率且对速度敏感的图像编码场景。
为高效压缩图像符号信息，本文研究了块级离散余弦变换（DCT）中的符号恢复问题：从系数幅度中重建DCT系数的符号。提出一种基于二分类机器学习的快速符号恢复方法。首先将幅度与符号构建为3D表示，将同频带的系数组织为2D子带块；随后通过卷积神经网络对3D幅度进行二分类，每个符号作为二元标签。实验表明，该方法在多个数据集上实现了接近100%的符号恢复准确率，同时计算开销极低，显著优于传统方法。
原文摘要 · Abstract (English)
To efficiently compress the sign information of images, we address a sign retrieval problem for the block-wise discrete cosine transformation (DCT): reconstruction of the signs of DCT coefficients from their amplitudes. To this end, we propose a fast sign retrieval method on the basis of binary classification machine learning. We first introduce 3D representations of the amplitudes and signs, where we pack amplitudes/signs belonging to the same frequency band into a 2D slice, referred to as the sub-band block. We then retrieve the signs from the 3D amplitudes via binary classification, where each sign is regarded as a binary label. We implement a binary classification algorithm using convolutional neural networks, which are advantageous for efficiently extracting features in the 3D amplitudes. Experimental results demonstrate that our method achieves accurate sign retrieval with an overwhelmingly low computation cost.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。