改进残差网络下体操动作逆光图像超分辨增强  

Super-Resolution Enhancement of Gymnastic Action Backlighting Image Based on Improved Residual N

作　　者：黄波[1] 赵新辉[1] HUANG Bo;ZHAO Xin-hui(Physical Education College,Zhengzhou University,Zhengzhou Henan 450044,China)

机构地区：[1]郑州大学体育学院,河南郑州450044

出　　处：《计算机仿真》2024年第5期231-235,309,共6页Computer Simulation

基　　金：国家重点研发计划重点专项项目(2020YFC2006800);河南省科技攻关项目(232102320309);河南省哲学社会科学规划项目(2022BTY022);河南省高等学校重点科研项目(22A890011)。

摘　　要：超分辨图像有利于信息获取与传递,但当下缺少对逆光低分辨率图像的研究。为解决逆光图像中阴影噪声的问题,提出一种基于改进Retinex逆光图像分解增光算法,通过结合DRAN双重残差注意力网络,构建出逆光图像超分辨率增强模型,即NRE-DRAN模型。模型首先将基于Retinex理论将逆光图像分解成光照量图与反射量图,并通过调节网络与恢复网络处理两分量图,然后将处理后的分量图融合,降低逆光噪声的影响;接着采用DRAN特征提取模块提取浅层纹理特征与深层语义特征,然后利用信息蒸馏模块增强特征信息,并串联拼接成融合特征图,最后基于残差图与上采样图重构出超分辨图像。多组基线算法叠加模型仿真对比结果表明,在GBI体操动作逆光图像数据集上,NRE-DRAN模型具有最优的SSIM指标(较其它叠加模型相比平均提升了3.93%)与较优的PSNR指标(指标排名第二)。综上所示,NRE-DRAN逆光图像超分辨率增强模型在解决逆光阴影噪声问题的同时有效的增强了图像的超分辨率,且该模型具有较高的时效性。Super-resolution images are beneficial to information acquisition and transmission,but there is a lack of research on backlighting low-resolution images.In order to solve the problem of shadow noise in backlighting images,this paper proposes a backlighting image super-resolution enhancement model based on an improved Retinex backlighting image decomposition and enhancement algorithm,which is called NRE-DRAN model by combining the DRAN double residual attention network.Firstly,the model decomposes the backlighting image into an illumination map and a reflection map based on Retinex theory,and processes the two-component maps by adjusting the network and restoring the network,and then fuses the processed component maps to reduce the influence of backlighting noise.Then,a DRAN feature extraction module is used to extract the shallow texture features and deep semantic features,and an information distillation module is used to enhance the feature information and concatenate them into a fused feature map.Finally,a super-resolution image is reconstructed based on the residual map and the up-sampled map.The simulation results of the multi-group baseline algorithm superposition model show that the NRE-DRAN model has the best SSIM index(an average increase of 3.93%compared with other superposition models)and the better PSNR index(the index ranks second)on the GBI gymnastics action backlighting image data set.To sum up,the NRE-DRAN backlighting image super-resolution enhancement model effectively enhances the super-resolution of the image while solving the problem of backlighting shadow noise,and the model has a high timeline.

关 键 词：超分辨增强 逆光图像 残差网络 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]