复合残差网络在低照度图像增强中的技术研究

Technical research of composite residual network in low illumination image enhancement

成像设备在暗光照环境下会出现对比度不高、图像细节信息丢失、颜色失真等问题,这会对视频监控、智能交通、人脸识别等应用场景产生巨大干扰。为了解决这一问题,本文提出了一种融合了注意力机制的的复合残差网络来实现对低照度图像的增强。该算法首先通过色彩空间上的转换(RGB-HSV)将亮度分量V放入构造的神经网络中,然后神经网络通过融合了注意力机制的多分支结构进行图像浅层特征的提取,接着经过复合残差网络提取深层特征,再经过图像重建得到增强后的V分量,最后通过分量融合实现图像增强。实验结果表明,对比目前国内外主流低照度图像增强算法,所提算法在主观视觉上对图像亮度与对比度有显著提升,在PSNR、SSIM指标上与传统算法的对比结果分别提升了约20%和15%,与深度学习算法的对比结果分别提升约9%和3%,不论是在人工合成的低照度图像还是真实、自然低照度图像中均有良好表现,基本满足图像增强的颜色自然、对比度和鲁棒性高等要求。

Imaging equipment in a dark environment has problems such as low contrast,loss of image detail information,and color distortion,which can cause huge interference in application scenarios such as video surveillance,intelligent transportation,and face recognition.In order to solve this situation,this paper proposes a composite residual network that incorporates the attention mechanism to enhance low-illuminance images.The algorithm firstly puts the brightness component V into the constructed neural network through color space conversion(RGB-HSV).The neural network extracts the shallow features of the image through a multi-branch structure that incorporates the attention mechanism,passes through the composite residual network extracts deep features,and then reconstructs the image to obtain the enhanced V component.Finally,the low-illuminance image enhancement is achieved through component fusion.The experimental results show that compared with the current mainstream low illumination image enhancement algorithms at home and abroad,the proposed algorithm significantly improves the image brightness and contrast in subjective vision.Compared with the traditional algorithm,the PSNR and SSIM are improved about 20%and 15%,respectively.And compared with the deep learning algorithm,the PSNR and SSIM are improved about 9%and 3%,respectively.It performs well in artificially synthesized low-light images or real and natural low-light images,basically meet the requirements of natural color,contrast and robustness for image enhancement.