基于含噪Retinex模型的煤矿低光照图像增强方法

A method for enhancing low light images in coal mines based on Retinex model containing noise

  • 摘要: 低光照图像会导致许多计算机 视觉任务达不到预期效果,影响后续图像分析与智能决策。针对现有煤矿井下低光照图像增强方法未考虑图像现实噪声的问题,提出一种基于含噪Retinex模型的煤矿低光照图像增强方法。建立了含噪Retienx模型,利用噪声估计模块(NEM)估计现实噪声,将原图像和估计噪声作为光照分量估计模块(IEM)和反射分量估计模块(REM)的输入,生成光照分量与反射分量并对二者进行耦合,同时对光照分量进行伽马校正等调整,对耦合后的图像及调整后的光照分量进行除法运算,得到最终的增强图像。NEM通过3层CNN对含噪图像进行拜耳采样,然后重构生成与原图像大小一致的三通道特征图。IEM与REM均以ResNet−34作为图像特征提取网络,引入多尺度非对称卷积与注意力模块(MACAM),以增强网络的细节过滤能力及重要特征筛选能力。定性和定量评估结果表明,该方法能够平衡光源与黑暗环境之间的关系,降低现实噪声的影响,在图像自然度、真实度、对比度、结构等方面均具有良好性能,图像增强效果优于Retinex−Net,Zero−DCE,DRBN,DSLR,TBEFN,RUAS等模型。通过消融实验验证了NEM与MACAM的有效性。

     

    Abstract: The low light images can lead to many computer vision tasks not achieving the expected results. This can affect subsequent image analysis and intelligent decision-making. The existing low light image enhancement methods for underground coal mines do not consider the real noise of the image. In order to solve this problem, a method for enhancing low light images in coal mines based on Retinex model containing noise is proposed. The Retienx model containing noise is established. The noise estimation module (NEM) is used to estimate real noise. The original image and estimated noise are used as inputs to the illumination component estimation module (IEM) and reflection estimation module (REM) to generate and couple the illumination and reflection components. At the same time, gamma correction and other adjustments are made to the illumination components. And division operations are performed on the coupled image and adjusted illumination components to obtain the final enhanced image. NEM uses a three-layer CNN to perform Bayer sampling on noisy images. It reconstructs them to generate a three channel feature map which is the same size as the original image. Both IEM and REM use ResNet-34 as the image feature extraction network. The multi-scale asymmetric convolution and attention module (MACAM) is introduced to enhance the network's capability to filter details and important features. The qualitative and quantitative evaluation results indicate that this method can balance the relationship between light sources and dark environments, reduce real-world noise's impact, and perform well in image naturalness, realism, contrast, structure, and other aspects. The image enhancement effect is superior to models such as Retinex-Net, Zero-DCE, DRBN, DSLR, TBEFN, RUAS, etc. The effectiveness of NEM and MACAM is verified through ablation experiments.

     

/

返回文章
返回