深度网络下的户外激光机器人视觉去雾方法

Outdoor Laser Robot Vision Defogging Method Under Deep Network

针对户外激光机器人视觉系统所捕获图像,由于大气中悬浮微粒所引起的色偏、失真、噪声等低能见度现象,提出一种深度网络去雾方法以提升户外激光机器人视觉感知能力。首先,在梯度域与灰度域中寻找平坦且连通的区域标定成天空光域,并采用四叉树方法获得大气光值;进而根据所输入的RGB特征图,并结合自编码网络提取纹理特征,完成非线性映射与透射图重建;最后,结合估计的大气光值与透射率代入大气散射模型复原清晰化图像。完成所提方法和retinex方法、双阈值分割法、自编码方法、AFF-Net方法的主客观对比实验,结果表明,所提方法能获得较好的全参考与无参考客观指标。因此,所提的户外激光机器人视觉去雾方法可以保持真实场景的原始色调,并生成细节丰富的无雾图像。

This paper proposes a deep network-based defogging method to enhance the outdoor laser robot's vision perception capability,addressing the low visibility issues caused by atmospheric suspended particles,such as color bias,distortion,noise,etc.,in the images.Firstly,flat and connected areas are identified and labeled as the sky light domain in the gradient domain and the gray domain,and the atmospheric light value is obtained using the quadtree method.Then,based on the input RGB feature map and the texture features extracted by the auto-encoder network,nonlinear mapping and transmission map reconstruction are completed.Finally,the estimated atmospheric light value and transmission rate are combined to restore the clear image using the atmospheric scattering model.Objective and subjective comparative experiments were conducted on our method,Retinex method,dual-threshold segmentation method,auto-encoder method,and AFF-Net method.The results showed that our method achieves superior objective indicators,both with and without reference,effectively preserving the original color tone of the real scene and generating clear,detailed images.