School of Computer Science and Technology, Henan Polytechnic University

针对目前井下尘雾图像清晰化算法存在的图像偏暗、细节丢失和过度增强等问题，提出一种基于增强网格网络的井下尘雾图像清晰化算法。该算法由前处理模块、主干模块和输出模块3个部分组成。前处理模块通过特征提取模块IRDB生成一组特征图，作为主干模块的输入，IRDB融合了Inception架构和密集残差连接模块（RDB）的优势，可在网络资源有限的情况下增加网络的深度和宽度，从而增强网络的表征能力、泛化能力及其对不同尺度尘雾的处理能力；主干模块采用网格网络进一步提取图像不同尺度的特征，并通过上采样和下采样实现特征图不同尺度的变换，为更好地捕捉图像中的细节信息，在网格网络中引入通道注意力机制。实验结果表明：IRDB数量为5时，网络模型的峰值信噪比（PSNR）、结构相似度指数（SSIM）和自然图像质量评价指标（NIQE）最好；从视觉效果上看，用所提算法清晰化处理后的图像细节信息更加丰富，色彩更加自然，具有良好的清晰度和对比度；在井下数据集上用所提算法处理后的图像PSNR、SSIM和NIQE分别为23.69，0.8401，8.95，图像处理速度处于中等水平，整体性能优于DCP，AOD−Net等同类算法。

To address the issues of dark images, detail loss, and over-enhancement in existing underground dust and mist image clarification algorithms, an image clarification algorithm based on enhanced grid networks was proposed. This algorithm consisted of three parts: a preprocessing module, a backbone module, and an output module. The preprocessing module generated a set of feature maps using the feature extraction module IRDB, which served as the input for the backbone module. The IRDB integrated the advantages of the Inception architecture and the Residual Dense Block (RDB), increasing the depth and width of the network under limited resources, thereby enhancing the network's representational ability, generalization capability, and handling of dust and mist at different scales. The backbone module employed a grid network to further extract features at various scales of the image and implemented transformations of feature maps at different scales through upsampling and downsampling. To better capture detailed information in the images, a channel attention mechanism was introduced within the grid network. Experimental results indicated that with 5 IRDB modules, the network model achieved the best Peak Signal to Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Naturalness Image Quality Evaluator (NIQE) metrics. Visually, the images processed using the proposed algorithm exhibited richer detail information, more natural colors, and improved clarity and contrast. The PSNR, SSIM, and NIQE values for the images processed by the proposed algorithm on the underground dataset were 23.69, 0.8401, and 8.95, respectively, with a moderate image processing speed, and the overall performance surpassed similar algorithms such as DCP and AOD-Net.

国家自然科学基金项目（52174174）；河南省高等学校重点科研项目（23B520002）；河南理工大学基本科研业务费专项项目（自然科学类）（NSFRF230429)。

谷亚楠，李晴，刘晨晨，等. 基于增强网格网络的井下尘雾图像清晰化算法[J]. 工矿自动化，2024，50（10）：120-127, 159.

GU Yanan, LI Qing, LIU Chenchen, et al. Image clarification algorithm for underground dust and mist based on enhanced grid network[J]. Journal of Mine Automation，2024，50（10）：120-127, 159.