Chinese Journal of Liquid Crystals and Displays, Volume. 39, Issue 8, 1024(2024)
HDR image processing algorithm for portrait based on multi feature fusion
Fig. 7. Comparison of effects before and after adding skin masks.（a）Input images；（b）Before adding skin mask；（c）After adding skin mask.
Fig. 8. Comparison of details before and after adding skin mask.（a）Before adding skin mask；（b）After adding skin mask.
Fig. 9. Histogram comparison.（a）Ground truth；（b）Adding the skin feature；（c）No adding the skin feature.
Fig. 10. Comparison of different algorithms on the PortraitHDR dataset.（a）Input images；（b）Ideal output images；（c）FHSVNet algorithm outputs images；（d）HDRnet algorithm outputs images；（e）U2HDRnet algorithm outputs images.
Table 1. Number of branch channels for different feature extraction
View tableView in ArticleTable 1. Number of branch channels for different feature extraction
支路 局部特征支路 全局特征支路 语义信息支路 Channels 64 256 512
Table 2. Comparison of different feature extraction networks
View tableView in ArticleTable 2. Comparison of different feature extraction networks
策略 特征提取网络改进策略 PSNR 语义特征 局部特征 全局特征动物 Gharbi √ √ 28.42 Ours（Bilateral） √ √ 29.23 Ours（Bilateral） √ √ 29.94 Ours（Bilateral） √ √ 29.98 Ours（Trilatera） √ √ √ 30.73
Table 3. Comparison of effects of different attention modules on different brands
View tableView in ArticleTable 3. Comparison of effects of different attention modules on different brands
注意力模块 PSNR 局部信息支路 全局信息支路 语义信息支路 特征融合支路 GACmix 30.68 30.51 30.75 31.36 Dual attention 29.99 30.27 30.72 31.04 NA 29.98 30.42 29.62 31.26 Channel Attention 30.49 30.28 30.37 31.21 SA 30.08 30.25 30.52 30.77 ACmix 30.06 30.14 30.21 31.03
Table 4. Comparison of effects before and after adding different skin segmentation module
View tableView in ArticleTable 4. Comparison of effects before and after adding different skin segmentation module
策略 SSIM No mask 0.915 +FCN 0.967 +Unet 0.964 +Unet3+ 0.968 +U2net 0.985
Table 5. Comparison of the effects of different fusion strategies
View tableView in ArticleTable 5. Comparison of the effects of different fusion strategies
策略 +三边特征提取模块 +注意力模块 +皮肤分割模块 PSNR SSIM 方案一 √ 30.73 0.909 方案二 √ √ 31.46 0.915 方案三 √ √ √ 31.42 0.985
Table 6. Comparison of the effects of different combination strategies
View tableView in ArticleTable 6. Comparison of the effects of different combination strategies
策略 局部特征 全局特征 语义特征 PSNR PSNR（+U2net） SSIM） SSIM（+U2net） 方案一 √ 27.32 27.29 0.866 0.916 方案二 √ 28.01 28.01 0.852 0.922 方案三 √ 27.98 27.96 0.867 0.928 方案四 √ √ √ 30.73 30.72 0.906 0.978
Table 7. Effects of different algorithms on the FiveK dataset
View tableView in ArticleTable 7. Effects of different algorithms on the FiveK dataset
算法 PSNR 风景 建筑 静物 车辆 动物 食品 HSVNet 26.52 23.35 25.22 25.41 22.18 26.37 HDRnet 28.96 29.65 28.22 27.47 25.14 28.01 Style-transfer 26.74 27.51 26.94 28.33 27.29 27.25 CNN-HDR 23.42 23.15 23.86 22.02 21.23 22.51 FHDR 29.93 30.52 30.39 30.85 27.58 29.86 Ours（no mask） 31.26 31.38 30.39 30.67 31.84 31.45
Table 8. Effect of different HDR algorithms in portrait scenes
View tableView in ArticleTable 8. Effect of different HDR algorithms in portrait scenes
算法 PSNR SSIM 训练时间/h 推理时间/s FHDR 28.32 0.829 20.6 0.24 HDRnet 28.38 0.904 16.8 0.10 Style-transfer 27.41 0.905 22.6 0.15 CNN-HDR 22.15 0.826 12.3 0.09 HSVNet 30.32 0.932 26.8 0.32 U2HDRnet 31.42 0.985 17.2 0.10
Get Citation
Copy Citation Text
Chunlin WU, Yongai ZHANG, Zhixian LIN, Tailiang GUO, Pengfei LIN, Jianpu LIN. HDR image processing algorithm for portrait based on multi feature fusion[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(8): 1024
Category: Image Enhancement
Received: Jul. 19, 2023
Accepted: --
Published Online: Sep. 27, 2024
The Author Email: Jianpu LIN (ljp@fzu.edu.cn)
