高速快门诱导的低照度图像弱参考视觉增强方法

高速快门会导致拍摄图像产生多种类型的退化,如极低曝光和噪声等问题.现有的无监督图像增强方法难以构建不同空间域的特征映射关系,以改善图像质量.针对上述问题,提出了一种高速快门诱导的低照度图像弱参考增强方法.该方法训练了一个光照特征提取网络(illumination feature extraction net,IFE-Net)以估计高阶曲线的参数;构建了联合硬注意力机制,加权选择低照度图像和参考图像的特征信息,并利用光照估计曲线将两者有机整合,逼近最佳的非线性映射,以获得清晰的复原图像;设计图像属性和转换感知相结合的多项损失函数,在增强低照度图像的同时保留更多图像细节.与现有的3种低照度图像增强算法进行实验对比,验证了算法的可行性和有效性,并通过消融实验验证了联合硬注意力模块设计的合理性和必要性.

High-speed shutter-induced weak reference enhancement method for low-illumination images

High-speed shutters can cause several types of degradation in captured images, such as problems with very low exposure and noise. Existing unsupervised image enhancement methods have difficulties in constructing feature mapping relationships in different spatial domains to improve image quality. To address the above problems, this paper proposes a high-speed shutter-induced weak reference enhancement method for low-illumination images. In this method, an illumination feature extraction net (IFE net) is trained to estimate the parameters of higher-order curves. A joint hard attention mechanism is constructed. The feature information of low illumination image and reference image is weighted, and the two are organically integrated by illumination estimation curve to approach the best nonlinear mapping to obtain a clear restored image. A multinomial loss function combining image attributes and transformation perception preserves more image details while enhancing the low-illumination image. Finally, experimental comparisons with three existing low-light image enhancement algorithms are conducted to verify the feasibility and effectiveness of the proposed method, and the rationality and necessity of the joint hard-attention module design is also validated through ablation experiments.