李宇娇, 周冬明, 李淼, 杨浩. 结合平滑扩展卷积网络与注意力机制的低照度图像增强[J]. 云南大学学报(自然科学版), 2022, 44(5): 912-924. doi: 10.7540/j.ynu.20220005
引用本文: 李宇娇, 周冬明, 李淼, 杨浩. 结合平滑扩展卷积网络与注意力机制的低照度图像增强[J]. 云南大学学报(自然科学版), 2022, 44(5): 912-924. doi: 10.7540/j.ynu.20220005
LI Yu-jiao, ZHOU Dong-ming, LI Miao, YANG Hao. Low-light image enhancement using smooth dilated convolutional network and attention mechanism[J]. Journal of Yunnan University: Natural Sciences Edition, 2022, 44(5): 912-924. DOI: 10.7540/j.ynu.20220005
Citation: LI Yu-jiao, ZHOU Dong-ming, LI Miao, YANG Hao. Low-light image enhancement using smooth dilated convolutional network and attention mechanism[J]. Journal of Yunnan University: Natural Sciences Edition, 2022, 44(5): 912-924. DOI: 10.7540/j.ynu.20220005

结合平滑扩展卷积网络与注意力机制的低照度图像增强

Low-light image enhancement using smooth dilated convolutional network and attention mechanism

  • 摘要: 近几年,卷积神经网络在低照度图像增强方向取得了显著成果. 然而,现有的大多数基于传统卷积神经网络的低照度增强模型效果受限于卷积中卷积核感受野是局部的,且池化层的应用使很多有价值的特征信息丢失. 为了解决这些问题,提出了一种端到端低照度图像增强网络. 首先,运用平滑扩展卷积层和Convolutional Block Attention Module(CBAM)注意力机制分别对图像进行特征提取;然后,通过拼接操作进行多层次特征融合;最后,多通道特征被送入到残差网络构成的重构网络,生成最终的增强图像. 此外,采用复合损失函数对低照度图像数据集进行训练. 实验结果表明,通过该方法,可以更好地实现图像多通道特征提取,主观视觉上提高了图像的亮度和对比度,更加符合人眼的视觉系统特性,与目前主流的图像增强方法相比,提出网络的平均客观图像质量指标PSNR、SNR在测试集上都优于其他算法.

     

    Abstract: In recent years, convolutional neural networks have made remarkable achievements in low light image enhancement. However, the effect of most existing low light enhancement models based on traditional convolutional neural networks is limited by the fact that the convolutional kernel sensory field is local in convolution, and the application of pooling layer makes a lot of valuable feature information lost. To solve these problems, we propose a lovel end-to-end low light image enhancement network. Firstly, the smooth dilated convolution layer and Convolutional Block Attention Module (CBAM) attention mechanism are used to extract the features of the image respectively. Secondly, the multi-layer feature fusion is carried out through the splicing operation. Finally, the multi-channel features are sent into the reconstruction network composed of residual network to generate the final enhanced image. In addition, we use compound loss function to train the low light image data set. Experimental results show that multi-channel feature extraction can be achieved better by using this method. Compared with the current mainstream image enhancement methods, the network enhancement effect proposed by us is superior to other methods in subjective vision and objective evaluation indicators.

     

/

返回文章
返回