arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2304.04416cs.CV
收藏

使用上下文感知 Transformer 的高动态范围成像

High Dynamic Range Imaging with Context-aware Transformer

Senslab Technology Co., Ltd.(森思实验室科技有限公司);NeuroSens Technology(神经感知科技公司)

Fangfang Zhou, Dan Zhang, Zhenming Fu

AI总结:

针对运动及饱和导致的 HDR 合成伪影,提出 HDT-HDR 分层双 Transformer,结合窗口注意力、可变形 CNN 与通道注意力同步建模全局和局部特征,实现最优去伪影效果。

摘要:

在将 LDR 图像合成为高动态范围(HDR)图像时避免引入伪影是一项具有挑战性的任务。卷积神经网络(CNN)通常对 HDR 去伪影有效,但如果 LDR 图像存在大幅运动或过饱和/欠饱和,则难以处理。现有的结合 CNN 与 Transformer 的双分支方法会遗漏非参考图像的部分信息,而基于 CNN 的分支所提取的特征受限于卷积核大小,感受野较小,这不利于去模糊以及过饱和/欠饱和区域的恢复。本文提出一种新颖的分层双 Transformer 方法,用于无伪影 HDR(HDT-HDR)图像生成,该方法可同时提取全局特征和局部特征。首先,我们使用带有空间注意力机制的基于 CNN 的头部从所有 LDR 图像中提取特征。其次,将 LDR 特征传递给分层双 Transformer(HDT)。在每个双 Transformer(DT)中,全局特征由基于窗口的 Transformer 提取,而局部细节则利用带有可变形 CNN 的通道注意力机制提取。最后,通过对 HDT 输出进行维度映射得到无伪影 HDR 图像。大量实验表明,我们的 HDT-HDR 在现有 HDR 去伪影方法中达到了最先进的性能。

英文摘要:

Avoiding the introduction of ghosts when synthesising LDR images as high dynamic range (HDR) images is a challenging task. Convolutional neural networks (CNNs) are effective for HDR ghost removal in general, but are challenging to deal with the LDR images if there are large movements or oversaturation/undersaturation. Existing dual-branch methods combining CNN and Transformer omit part of the information from non-reference images, while the features extracted by the CNN-based branch are bound to the kernel size with small receptive field, which are detrimental to the deblurring and the recovery of oversaturated/undersaturated regions. In this paper, we propose a novel hierarchical dual Transformer method for ghost-free HDR (HDT-HDR) images generation, which extracts global features and local features simultaneously. First, we use a CNN-based head with spatial attention mechanisms to extract features from all the LDR images. Second, the LDR features are delivered to the Hierarchical Dual Transformer (HDT). In each Dual Transformer (DT), the global features are extracted by the window-based Transformer, while the local details are extracted using the channel attention mechanism with deformable CNNs. Finally, the ghost free HDR image is obtained by dimensional mapping on the HDT output. Abundant experiments demonstrate that our HDT-HDR achieves the state-of-the-art performance among existing HDR ghost removal methods.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑