基于改进U-net多尺度注意力的多曝光图像融合算法

Multi-exposure image fusion algorithm based on improved U-net multi-scale attention

自然场景具有宽广的动态范围,但是现有的商用成像设备并不能达到较好记录这些高动态范围的需求,因此多曝光图像融合的提出是一种经济、快速且高效的高动态范围(HDR)实现方法。然而,现阶段的多曝光图像融合算法存在数据需求量大,融合结果存在对比度差和色彩失真等问题。针对这些问题,提出了一种基于多尺度注意力的多曝光图像融合方法,该方法将两幅极端曝光图像序列发送到网络,通过多尺度模块和通道、空间双注意力机制,自主学习并输出最优融合权重。此外,模型采用了真实值训练,并通过一个新的自定义损失函数使输出更接近真实图像。实验结果表明,该方法在客观和主观方面均优于现有方法。

Natural scenes have a wide dynamic range, but existing commercial imaging equipment does not meet the need to record these high dynamic ranges, so multi-exposure image fusion is proposed as an economical, fast and efficient way to implement high dynamic range(HDR) imaging. However, the current multi-exposure image fusion algorithm suffers from high data demand, poor contrast and color distortion of the fused images. To address these problems, a multi-exposure image fusion method based on multi-scale attention is proposed in this paper. The method sends two sequences of extreme exposure images to the network, and learns and outputs the optimal fusion weights autonomously through the multiscale module and the channel and spatial dual-attention mechanism. In addition, the model in this paper is trained with real values, and the output is made closer to the real image by a new custom loss function. The experimental results show that the method outperforms existing methods in both objective and subjective aspects.