矿井图像超分辨率重建研究

Research on super-resolution reconstruction of mine images

受井下粉尘大、照度低等环境影响,矿井图像存在分辨率低、细节模糊等问题,现有的图像超分辨率重建算法应用于矿井图像时,难以获取不同尺度图像信息、网络参数过大而影响重建速度,且重建图像易出现细节丢失、边缘轮廓模糊、伪影等问题。提出了一种基于多尺度密集通道注意力超分辨率生成对抗网络(SRGAN)的矿井图像超分辨率重建算法。设计了多尺度密集通道注意力残差块替代SRGAN原有的残差块,采用2路并行且卷积核大小不同的密集连接块,可充分获取图像特征;融入高效通道注意力模块,加强对高频信息的关注度;采用深度可分离卷积对网络进行轻量化,抑制网络参数的增加;利用纹理损失约束网络训练,避免网络加深时产生伪影。在井下数据集和公共数据集上对提出的矿井图像超分辨率重建算法和经典超分辨率重建算法BICUBIC,SRCNN,SRRESNET,SRGAN进行实验,结果表明:所提算法在主客观评价上总体优于对比算法,网络参数较SRGAN减少了2.54%,峰值信噪比与结构相似度较经典算法指标均值分别提高了0.764 dB和0.05358,能更好地关注图像的纹理、轮廓等细节信息,重建图像更符合人眼视觉。

Due to the impact of high dust and low illumination in underground environments,mine images have problems such as low resolution and blurry details.When existing image super-resolution reconstruction algorithms are applied to mine images,it is difficult to obtain image information at different scales.The network parameters are too large,which affects the reconstruction speed.The reconstructed images are prone to problems such as detail loss,blurry edge contours,and artifacts.A mine image super-resolution reconstruction algorithm based on multi-scale dense channel attention super-resolution generative adversarial network(SRGAN)is proposed.A multi-scale dense channel attention residual block is designed to replace the original residual block of SRGAN.Two parallel dense connected blocks with different convolutional kernel sizes are used to fully obtain image features.The efficient channel attention modules are integrated to enhance attention to high-frequency information.The depthwise separable convolution is used to lighten the network and suppress the increase of network parameters.The texture loss constraint network training is utilized to avoid artifacts during network deepening.Experiments are conducted on the proposed mine image super-resolution reconstruction algorithm and classic super-resolution reconstruction algorithms BICUBIC,SRCNN,SRRESNET,SRGAN on both underground and public datasets.The results show that the proposed algorithm outperformed the comparative algorithm in both subjective and objective evaluations.Compared to SRGAN,the proposed algorithm reduces network parameters by 2.54%.Compared to the average index values of the classic algorithms,the peak signal-tonoise ratio and structural similarity of the proposed algorithm increase by 0.764 dB and 0.05358 respectively.It can better focus on the texture,contour and other details of the image,and the reconstructed image is more in line with human vision.