激光与光电子学进展, 2020, 57 (20): 201004, 网络出版: 2020-10-14
基于注意力机制和Retinex的低照度图像增强方法 下载: 1232次
Low-Illumination Image Enhancement Method Based on Attention Mechanism and Retinex
图像处理 图像增强 低照度图像增强 Retinex算法 注意力机制 image processing image enhancement low-illumination image enhancement Retinex algorithm attention mechanism
摘要
低照度图像增强的主要目的是提升图像的整体光照度,进而呈现更多有用的信息。针对低照度图像的整体照度低、对比度弱和噪声较高的问题,提出基于注意力机制和Retinex算法的低照度图像增强方法。该方法首先将低照度图像分解为不变性反射图和缓变平滑性光照图;再通过注意力机制提取图像的空间和局部物体信息,从而能够保证增强过程中利用空间和局部物体信息进行约束;同时增加色彩损失函数改善图像饱和度,用以补偿和校准增强过程中的对比度细节;改进低照度图像和合成方法,加入真实噪声有效扩充训练数据集。最终在LOL和SID数据集上实验表明,所提方法的主观感受和客观评价指标均有所提升。
Abstract
The goal of low-illuminance image enhancement is to increase the overall illuminance of the image, thereby presenting more useful information. Aiming at the problems of low illumination, low contrast and high noise in low-illumination images, a method of image enhancement method based on attention mechanism and Retinex algorithm is proposed. This method first decomposes the low-illumintion image into an invariant reflection map and a slowly-varying smooth illumination map. Then, it uses the attention mechanism to extract the spatial and local object information of the image, so as to ensure that the spatial and local object information is used for constraints during the enhancement process. Moreover, it increases the color loss function to improve the image saturation to compensate and calibrate the contrast details in the enhancement process. Furthermore, it improves the low-illumintion image and synthesis method, add real noise, and efficiently expands the training data set. Finally, the experiments on the LOL and SID data sets show that the subjective and objective evaluation indicators of the proposed method improved.
黄辉先, 陈凡浩. 基于注意力机制和Retinex的低照度图像增强方法[J]. 激光与光电子学进展, 2020, 57(20): 201004. Huixian Huang, Fanhao Chen. Low-Illumination Image Enhancement Method Based on Attention Mechanism and Retinex[J]. Laser & Optoelectronics Progress, 2020, 57(20): 201004.