基于多分支全卷积神经网络的低照度图像增强 下载: 958次
Low-Light Image Enhancement Based on Multi-Branch All Convolutional Neural Network
图 & 表
图 1. 卷积块注意力模块结构
Fig. 1. Structure of convolutional block attention module
下载图片 查看原文
图 2. 多分支全卷积神经网络架构
Fig. 2. Architecture of multi-branch all convolutional neural network
下载图片 查看原文
图 3. 噪声提取模块
Fig. 3. Noise extraction module
下载图片 查看原文
图 4. 不含噪声的合成低照度图像的主观视觉对比。(a) Parrots图像;(b) building2图像;(c) buildings图像;(d) monarch图像
Fig. 4. Subjective visual comparison of synthetic low-light images without noise. (a) Image of parrots; (b) image of building2; (c) image of buildings; (d) image of monarch
下载图片 查看原文
图 5. 含噪声的合成低照度图像的主观视觉对比。(a) Parrots图像;(b) building2图像;(c) buildings图像;(d) monarch图像
Fig. 5. Subjective visual comparison of synthetic low-light images with noise. (a) Image of parrots; (b) image of building2; (c) image of buildings; (d) image of monarch
下载图片 查看原文
图 6. 真实低照度图像的主观视觉对比。(a)(b)来自LIME数据库的图像;(c)(d)来自DICM数据库的图像;(e)来自MEF数据库的图像
Fig. 6. Subjective visual comparison of real low-light images. (a)(b) Images from LIME dataset; (c)(d) images from DICM dataset; (e) image from MEF dataset
下载图片 查看原文
表 1不含噪声的合成低照度图像的客观评价指标对比
Table1. Comparison of objective evaluation indicators for synthetic low-light images without noise
Image | PSNR/dB | MSE | MAE | MS-SSIM | Q | VIF |
---|
Originalimage | 10.036/9.460 | 6447.9000/7440.8000 | 74.8505/76.8100 | 0.3925/0.3400 | 0.2407/0.2400 | 0.4889/0.5100 | Image ofCLAHE[3] | 15.596/15.850 | 1792.4000/1817.0000 | 39.3139/36.4600 | 0.7988/0.7815 | 0.6357/0.6900 | 0.7582/0.7500 | Image ofSSR[6] | 22.389/20.560 | 375.1150/658.8000 | 14.3500/19.1700 | 0.8247/0.7800 | 0.6489/0.6800 | 0.8437/0.7400 | Image ofMSRCR[8] | 11.135/11.410 | 5006.9000/4775.5000 | 67.8155/62.9100 | 0.7602/0.6900 | 0.6030/0.5900 | 0.6348/0.5800 | Image of methodin Ref. [12] | 19.040/17.430 | 811.0260/1220.0000 | 22.8847/28.7300 | 0.8001/0.7400 | 0.6243/0.6300 | 0.6229/0.4200 | Image of methodin Ref. [30] | 19.959/18.250 | 656.4077/1028.0000 | 18.3382/22.6600 | 0.8439/0.7800 | 0.6522/0.6500 | 0.6801/0.6700 | Image of methodin Ref. [16] | 22.679/21.860 | 350.8300/434.6700 | 16.5200/17.3000 | 0.9130/0.8650 | 0.7360/0.7640 | 0.8750/0.8510 | Image ofMBACNN | 23.869/22.550 | 266.7760/384.2000 | 13.8820/15.5000 | 0.9229/0.8700 | 0.7735/0.7700 | 0.8834/0.8630 |
|
查看原文
表 2含噪声的合成低照度图像上的客观评价指标对比
Table2. Comparison of objective evaluation indicators for synthetic low-light images with noise
Image | PSNR/dB | MSE | MAE | MS-SSIM | Q | VIF |
---|
Originalimage | 9.8380/8.7000 | 6749.6000/8878.0000 | 74.8583/84.9000 | 0.3050/0.2200 | 0.1364/0.0800 | 0.3249/0.3400 | Image ofCLAHE[3] | 13.4239/11.9600 | 2955.9000/4215.0000 | 49.0863/56.7000 | 0.5606/0.4100 | 0.3990/0.2700 | 0.5170/0.5100 | Image ofSSR[6] | 18.9663/18.3000 | 825.0010/979.0000 | 21.3554/23.6000 | 0.6690/0.5100 | 0.5018/0.4000 | 0.6058/0.5800 | Image ofMSRCR[8] | 10.4440/17.0400 | 5870.6000/1344.0000 | 72.3333/28.9000 | 0.5592/0.5000 | 0.3809/0.3750 | 0.5436/0.6280 | Image of methodin Ref. [12] | 19.5938/16.2000 | 714.0061/1600.0000 | 19.2488/33.7000 | 0.6630/0.4800 | 0.4639/0.4100 | 0.5216/0.4200 | Image of methodin Ref. [30] | 17.8858/16.5000 | 1058.0000/1535.7000 | 22.6885/30.3000 | 0.6145/0.5400 | 0.3896/0.3200 | 0.4593/0.4700 | Image of methodin Ref. [16] | 18.3000/19.3300 | 961.3000/776.5500 | 25.1700/21.6700 | 0.6900/0.6200 | 0.5300/0.4800 | 0.7370/0.7190 | Image ofMBACNN | 21.1500/19.8200 | 499.0299/697.3000 | 16.3177/20.2400 | 0.7970/0.6800 | 0.6251/0.5200 | 0.6903/0.7400 |
|
查看原文
表 3真实低照度图像的客观评价指标对比
Table3. Comparison of objective evaluation indicators for real low-light images
Image | NRSS | Entropy of information | NIQE |
---|
Image of CLAHE[3] | 0.8862/0.9500 | 7.6180/7.3000 | 3.9221/5.2000 | Image of SSR[6] | 0.9257/0.9400 | 7.6845/7.3600 | 4.2682/5.0100 | Image of MSRCR[8] | 0.8946/0.9300 | 7.3285/7.5300 | 3.9577/4.2600 | Image of method in Ref. [12] | 0.9159/0.9190 | 7.7841/7.4600 | 4.1566/4.6100 | Image of method in Ref. [30] | 0.8785/0.9220 | 7.2858/7.4180 | 4.4679/4.9800 | Image of method in Ref. [16] | 0.9240/0.9630 | 6.6100/6.2250 | 5.1130/5.2530 | Image of MBACNN | 0.9359/0.9790 | 7.5844/7.4900 | 5.3669/4.6770 |
|
查看原文
表 4算法处理时间对比
Table4. Algorithm processing time comparison
Method | Training time /h | Test time /s |
---|
CLAHE[3] | — | 0.77 | SSR[6] | — | 2.65 | MSRCR[8] | — | 1.02 | Method in Ref. [12] | — | 0.72 | Method in Ref. [30] | — | 1.66 | Method in Ref. [16] | 8.96 | 5.88 | MBACNN | 5.99 | 6.78 |
|
查看原文
吴若有, 王德兴, 袁红春, 宫鹏, 陈冠奇, 王丹. 基于多分支全卷积神经网络的低照度图像增强[J]. 激光与光电子学进展, 2020, 57(14): 141021. Ruoyou Wu, Dexing Wang, Hongchun Yuan, Peng Gong, Guanqi Chen, Dan Wang. Low-Light Image Enhancement Based on Multi-Branch All Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141021.