融合层次特征和注意力机制的轻量化矿井图像超分辨率重建方法

Lightweight super-resolution reconstruction method based on hierarchical features fusion and attention mechanism for mine image

针对矿井图像灰暗模糊、边缘不清晰等问题,提出了一种融合层次特征和注意力机制的轻量化矿井图像超分辨率重建方法。首先设计一种残差坐标注意力模块,在残差块中融入坐标注意力机制,使网络获得更丰富的高频细节信息;其次采用层次特征融合机制,对不同网络层次的特征信息进行特征融合,促进边缘细节信息的重建。最后,再对融合后的特征进行降维以减少模型计算量和参数量。为了使模型在真实矿井场景中具有更好的泛化能力,构建了一种煤矿井下图像数据集CMUID用于网络模型的训练和测试实验。实验结果表明,本文算法的重建图像质量在客观指标和主观感受上均优于其他对比算法。当缩放因子为4时,与OISR算法相比,在煤矿井下数据集上PSNR和SSIM的平均值分别提升了0.3185 dB和0.0126,在公共数据集上PSNR和SSIM的平均值分别提升了0.1 dB和0.0035;网络模型参数量减少了70.7%。

The images in coal mines have problems of dim,blurry and unclear edges.To address these issues,this article proposes a lightweight mine image super-resolution reconstruction method that fuses hierarchical features and attention mechanism.Firstly,by integrating the coordinate attention mechanism into the residual block,this article designs a residual coordinate attention module,which enables the network to obtain rich high-frequency detailed information.Secondly,the hierarchical feature fusion mechanism is adopted to fuse the feature map information of different network levels.Thereby,the reconstruction of edge detail information is promoted.Finally,the dimensionality reduction is performed on the fused features to reduce the amount of model computation and parameters.In addition,to make the proposed model have better generalization performance in real-mine scenes,a coal mine underground image dataset CMUID is constructed for the training and testing experiments of the network model.Experimental results demonstrate that the reconstructed image quality of the proposed algorithm is superior to other comparison algorithms in both objective indicators and subjective feelings.Compared with the OISR algorithm on the underground coal mine image data set,when the scaling factor is set to 4,the average values of PSNR and SSIM of the proposed algorithm can be improved by 0.3185 dB and 0.0126.As for the public data set,the average PSNR and SSIM of the proposed algorithm are also improved by 0.1 dB and 0.0035,respectively,as well as the number of network model parameters is reduced by 70.7%.