| 摘要: |
| 目的 低光工况或拍摄技巧影响都可能获得低光照图像,为解决此类图像对比度低、噪声大、颜色失真等问
题,提出一种卷积神经网络增强模型 RetKIND,包括分解网络、亮度调整网络和降噪网络。 方法 它借助残差模块
(RB)和跳跃连接,有效抑制分解网络在分解时产生的噪声;融合 U-Net 架构、空洞卷积和 EBAM 高效注意力机制
构建降噪网络,利用空洞卷积扩大感受野,提取更多图像信息,提高 EBAM 在通道和空间上提取反射图的细节、纹
理、颜色等特征的能力,实现图像去噪;由 UC(亮度调整网络中的自定义模块)和普通卷积组成亮度调整网络,旨在
减少光照图细节缺失,提高光照分量对比度。 融合去噪后的反射分量和增强后的光照分量,得到正常光照图像。
结果 仿真结果表明:在 LOL 数据集上,相较 R2RNet,FPSNR 和 FSSIM 值分别上升了 6. 2%和 4. 2%;相较 URetinexNet,FPSNR 和 FSSIM 值分别上升了 5. 9%和 1. 2%;相较 DEANet,FPSNR 和 FSSIM 值分别上升了 2. 9%和 1. 1%。 结论 Ret
-KIND 模型既能提升图像亮度,又能降低图像的噪声,有助于推动低光图像增强模型应用到目标检测领域。 |
| 关键词: 低光照图像增强 去噪 RB 残差模块 EBAM 注意力机制 Retinex 理论 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Low-light Image Enhancement Algorithm Integrating Residual Structure and Attention Mechanism |
|
LIU Yao；JIA Xiaofen
|
|
School of Electrical and Information Engineering Anhui University of Science & Technology Anhui Huainan 232001
China
|
| Abstract: |
| Objective Low-light images can be obtained due to either low-light conditions or shooting techniques. In orde
to solve such problems as low contrast high noise and color distortion of images a convolutional neural network
enhancement model RetKIND was proposed which included a decomposition network brightness adjustment network
and noise reduction network. Methods With the help of the residual block RB and skip connection the noise
generated by the decomposition network during decomposition was effectively suppressed. The noise reduction network was
constructed by integrating U-Net architecture dilated convolution and EBAM efficient attention mechanism. The dilated
convolution was used to enlarge the receptive field to extract more image information and EBAM was utilized to extract
details texture color and other features of the reflection image in channel and space to achieve image denoising. The
brightness adjustment network was composed of a UC module self-designed module in the brightness adjustment network
and traditional convolution which aimed to reduce the detail loss of light images and improve the contrast of the light
component. Finally the denoised reflection component and the enhanced light component were fused to obtain the normal
illumination image. Results Simulation results showed that on the dataset LOL compared with R2RNet the values of
FPSNR and FSSIM increased by 6. 2% and 4. 2% respectively compared with URetinex-Net the values of FPSNR
and FSSIM of Ret-KIND increased by 5. 9% and 1. 2% respectively compared with DEANet the values of FPSNR
and FSSIM of RetKIND increased by 2. 9% and 1. 1% respectively. Conclusion The Ret-KIND model can not only improve image
brightness but also reduce image noise which helps to promote the application of the low-light image enhancement model
to the field of target detection. |
| Key words: low-light image enhancement denoising RB residual module EBAM attention mechanism Retinex theory |
