摘要
煤矿井下环境复杂,尘雾干扰易导致图像质量下降,影响后续分析。循环生成对抗网络(CycleGAN)算法在井下尘雾图像处理中具有数据获取和训练优势,但存在细节丢失、色彩失真及泛化能力不足等问题。基于此,提出了基于CycleGAN改进的CM-CycleGAN算法,通过优化生成器和判别器结构,引入特征增强模块以提升多尺度尘雾特征提取能力,并结合注意力机制强化关键区域特征学习。同时,改进损失函数,融合循环感知损失和颜色损失,在提升去雾效果的同时保持图像色彩与结构特征。实验表明,CM-CycleGAN在峰值信噪比(PSNR)、结构相似性指数(SSIM)和雾密度估计(FADE)指标上分别平均提升1.8300、0.0141和降低0.0415,显著优于对比方法,为煤矿井下尘雾图像处理提供了有效解决方案。
The coal mine underground environment is complex,where dust and haze interference significantly degrade image quality and hinder subsequent analysis.While the cycle generative adversarial networks(CycleGAN)algorithm demonstrates advantages in data acquisition and model training for underground dust-haze image processing,it suffers from issues such as loss of fine details,color distortion,and weak generalization capability across diverse mining scenarios.To address these limitations,an enhanced CycleGAN-based framework named CM-CycleGAN was proposed for dust-haze removal in underground coal mine images.The proposed method optimized the generator and discriminator architectures by incorporating a feature enhancement module to improve multi-scale dust-haze feature extraction,and an attention mechanism was further integrated to emphasize critical features in contaminated regions.Additionally,the loss function was refined by combining cycle-consistency perception loss and color loss,preserving structural integrity and chromatic fidelity while enhancing dehazing effect in the generated images.Experimental results demonstrated that CM-CycleGAN outperformed conventional methods,achieving average improvements of 1.83 in peak signal-to-noise ratio(PSNR),0.0141 in structural similarity index(SSIM),and a reduction of 0.0415 in fog aware density evaluator(FADE)metrics.This study provides an effective technical solution for dust-haze image restoration in coal mines.
作者
汤璧屾
毛善君
智宁
吴峥
樊迎博
TANG Bishen;MAO Shanjun;ZHI Ning;WU Zheng;FAN Yingbo(School of Earth and Space Sciences,Peking University,Haidian,Beijing 100871,China;Ordos Research Institute of Energy,Peking University,Ordos,Inner Mongolia 017010,China;Institute of Electronics,Chinese Academy of Sciences,Haidian,Beijing 100190,China)
出处
《中国煤炭》
北大核心
2025年第12期163-172,共10页
China Coal
基金
国家重点研发计划项目“煤矿重大灾害演化与监控数字孪生技术研究”(2022YFC3004701)
鄂尔多斯市国家可持续发展议程创新示范区建设(含厅市会商、科技突围工程)科技支撑项目“高级智能化矿山构建关键技术研发及应用”(KCX2024002)。
关键词
尘雾图像清晰化
生成对抗网络
煤矿井下图像
CycleGAN
dust-haze image enhancement
generative adversarial network(GAN)
underground coal mine images
CycleGAN
