寇旗旗,程志威,程德强,等. 基于蓝图分离卷积的轻量化矿井图像超分辨率重建方法[J]. 煤炭学报,2024,49(9):4038−4050. DOI: 10.13225/j.cnki.jccs.2023.1101
引用本文: 寇旗旗,程志威,程德强,等. 基于蓝图分离卷积的轻量化矿井图像超分辨率重建方法[J]. 煤炭学报,2024,49(9):4038−4050. DOI: 10.13225/j.cnki.jccs.2023.1101
KOU Qiqi,CHENG Zhiwei,CHENG Deqiang,et al. Lightweight super resolution method based on blueprint separable convolution for mine image[J]. Journal of China Coal Society,2024,49(9):4038−4050. DOI: 10.13225/j.cnki.jccs.2023.1101
Citation: KOU Qiqi,CHENG Zhiwei,CHENG Deqiang,et al. Lightweight super resolution method based on blueprint separable convolution for mine image[J]. Journal of China Coal Society,2024,49(9):4038−4050. DOI: 10.13225/j.cnki.jccs.2023.1101

基于蓝图分离卷积的轻量化矿井图像超分辨率重建方法

Lightweight super resolution method based on blueprint separable convolution for mine image

  • 摘要: 煤矿井下复杂受限空间中,人造光源照明不均匀、工作面尘雾浓度大及复杂电磁干扰等环境因素都会严重影响井下监控视频的高清成像。针对受复杂环境影响下的矿井图像出现分辨率较低、模糊不清的问题以及目前图像超分辨率重建方法多以牺牲网络深度和宽度为代价来提高图像重建效果,从而导致算法的复杂度大幅增加、网络模型的内存占用率变大、难以应用到实际边缘移动设备中的问题,提出了一种基于蓝图分离卷积的轻量化矿井图像超分辨率重建方法。首先,该方法使用高效率的蓝图分离卷积替换残差块中的标准卷积,设计出一种轻量残差注意力模块,接着引入坐标注意力机制并添加跳跃连接对残差块进行改进,使模型保持较低参数量和计算量的同时也具备良好的特征提取能力。其次设计了一种增强层次特征融合模块,对网络中的不同层次特征以先局部再全局的方式融合,进一步促进网络中的信息流动,增强模型的特征利用率。最后在网络末端添加像素注意力机制,用于增强网络对信息特征的关注度,提高模型的特征表达能力,为图像重建模块提供更丰富的细节特征。实验结果表明,基于蓝图分离卷积的轻量化超分辨率重建网络所重建后的图像质量不仅在客观指标和视觉感受上均优于其他对比算法,而且能够在模型性能和复杂度之间取得更好的权衡。当缩放因子为4时,相比于轻量级算法AWSRN-M,参数量相当的情况下,在煤矿井下测试集上客观指标PSNR平均值提升了0.177 2 dB,SSIM平均值提升了0.010 7,浮点运算量减少了66.9%。结果证明了所提方法可以有效提取不同层次的细节特征信息,实现浅层特征和深层特征的深度融合,且更加高效地重建出纹理细节清晰的高分辨率图像。

     

    Abstract: In the complex confined space of underground coal mine, some environmental factors such as uneven illumination of artificial light source, heavy concentration of dust and fog in the working face, and complex electromagnetic interference can seriously affect the high-definition imaging of underground surveillance video. Aiming at the problems of low resolution and blurring of mine images affected by complex environment, and the problems that the current image super-resolution reconstruction methods mostly improve the reconstruction effect at the expense of compromising network depth and width, which will increase the complexity of the algorithm greatly, as well as the memory usage of the network model, making it difficult to apply to actual edge mobile devices. In this study, a lightweight super-resolution reconstruction method of mine images based on blueprint separation convolution is proposed. Firstly, the high-efficiency blueprint separation convolution is used to replace the standard convolution in the residual block, and a lightweight residual attention module is designed by introducing coordinate attention mechanism and adding skip connections to improve the residual block, so that the model can achieve a better feature extraction ability while keeping low parameters and computation. Secondly, an enhanced hierarchical feature fusion module is designed to integrate local features first and then global features of different hierarchical features in the network, which can further promote the information flow in the network and enhance the feature utilization rate of the model. Finally, the pixel attention mechanism is added at the end of the network to enhance the attention of the network to the information features, which can improve the feature expression ability of the model and provide more detailed features for the image reconstruction module. Experimental results show that the image quality reconstructed by the lightweight super-resolution reconstruction network based on blueprint separation convolution is not only superior to other comparison algorithms in terms of objective indexes and visual perception, but also achieves a better trade-off between model performance and complexity. When the scaling factor is 4, in comparison to the lightweight algorithm AWSRN-M with the same number of parameters, the average PSNR and SSIM on the coal mine test set are increased by 0.177 2 dB and 0.010 7 respectively, and the floating-point computation is reduced by 66.9%. The results demonstrate that the proposed method can effectively extract the detailed feature information with different levels, achieve the deep fusion of shallow and deep features, and more efficiently reconstruct high-resolution images with clear texture details.

     

/

返回文章
返回