基于改进CycleGAN的煤矿井下低照度图像增强方法  

Image enhancement method of underground low illumination in coal mine based on improved CycleGAN

作　　者：吴佳奇 张文琪 陈伟[1,2] 王帅[1,3] WU Jiaqi;ZHANG Wenqi;CHEN Wei;WANG Shuai(School of Mechanical Electronic and Information Engineering,China University of Mining and Technology(Beijing),Beijing 100083,China;School of Computer Science and Technology,China University of Mining and Technology,Xuzhou 221116,Jiangsu China;Department of Inner Mongolia Administration of Coal Mine Safety Ordos Division,Ordos 017000,Inner Mongolia China)

机构地区：[1]中国矿业大学(北京)机电与信息工程学院,北京100083 [2]中国矿业大学计算机科学与技术学院,江苏徐州221116 [3]内蒙古煤矿安全监察局鄂尔多斯监察分局,内蒙古自治区鄂尔多斯017000

出　　处：《华中科技大学学报（自然科学版）》2023年第5期40-46,共7页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(52274160,52074305,51874300);国家自然科学基金委员会-山西省人民政府煤基低碳联合基金资助项目(U1510115).

摘　　要：针对井下监控装置采集的图像普遍存在照度低、颜色失真及细节特征损失严重等缺陷,提出一种基于改进CycleGAN网络的煤矿井下低照度图像增强算法.首先针对井下成对图像数据获取困难的问题,基于循环生成对抗网络搭建循环图像增强主体框架实现模型的无监督训练;然后基于CSDNet的全局图像分解架构,设计了一种融合空间-通道注意力模块CBAM的双分支估计网络以并行估计图像的光照分量和反射分量,并在两分支网络之间建立多尺度特征分解机制,从而在大幅提升亮度的同时避免颜色失真现象,保留大量细节信息;使用全局-局部判别器调节图像局部区域的亮度,改善亮度不均,避免过曝及阴影现象.实验结果表明:相较于对比算法RetinexNet,LLNet,MBLLEN,EnlightenGAN和CSDNet,本算法在客观质量指标PSNR(峰值信噪比)、SSIM(结构相似性)、IFC(信息保真度)和VIF(视觉信息保真度)上的表现分别提高了11.787%,8.256%,9.658%和8.654%,并在人类视觉主观分析上优于对比算法,证明本文算法能够有效改善井下低照度图像视觉效果.Aiming at the defects of low illumination,color distortion and serious loss of detail features in the images collected by the underground monitoring device,a image enhancement algorithm of underground low illumination in coal mine based on the improved CycleGAN network was proposed.First,aiming at the difficulty of acquiring underground paired image data,a cyclic image enhancement framework was built based on the cyclic generation antagonism network to realize unsupervised training of the model.Then,based on the image decomposition architecture of CSDNet,a dual-branch estimation network integrating the spacechannel attention module CBAM was designed to estimate the illumination and reflection components of the image in parallel,and a multi-scale feature decomposition mechanism was established between the two branch networks,so as to avoid color distortion and retain a large amount of detail information while greatly improving the brightness.The global-local discriminator was used to adjust the brightness,improve the uneven brightness,and avoid over-exposure and shadow of the local area of the image.Experiment results show that compared with the comparison algorithm RetinexNet,LLNet,MBLLEN,EnlightenGAN and CSDNet,the performances of the proposed algorithm on the objective quality indicators PSNR(peak signal to noise ratio),SSIM(structural similarity),IFC(information fidelity criterion)and VIF(visual information fidelity)are improved by 11.787%,8.256%,9.658%and 8.654%,respectively,and is superior to the comparison algorithm in the subjective analysis of human vision,which proves that the algorithm can effectively improve the visual effect of underground low illumination images.

关 键 词：井下低照度图像增强 颜色失真 无监督学习 RETINEX算法 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]