强化边缘特征提取的双分支融合真实图像去雾

打开文本图片集

Enhanced edge feature extraction dual branch fusion network for real image dehazing

LI Xiongxin，XIA Fengling，ZHANG Kaomin，WANG Hongliang，XIE Tao （Faculty of Civil Auiation and Aeronautics ，Kunming University of Science and Technology， Kunming 650500，China） * Corresponding author， E -mail： 77183118@qq. com

Abstract： Haze in natural environments is usually non-homogeneous and irregular，which has a large impact on computer vision tasks. Therefore，this paper proposed the Enhanced-edge-feature Dual-branch Fusion Dehazing Network （EDFDNet）. In order to retain the realism of the image and at the same time effectively improve the visibility after dehazing in the case of severe blurring，the transmission graph fine branch was constructed，which was the premier branch of the network. A U-shaped network hierarchical codec structure that fused the discrete wavelet transform was used to extract multi-scale fine feature information，and a mathematical method for the determination of the enhanced edge information was defined. The feature extraction branch tandemly connected the ResNet residual block and the Transformer combined with dual attention for a paralel feature extraction module，which fused the extracted local and glob al features.This improved the network's ability to understandand process non-uniform haze images and further restored the visibility of the images.These two branches were joined into the backbone framework of the Generative Adversarial Network （GAN），and a mathematical method to strengthen the determination ofedge information was defined. This formed the defogging network EDFDNet.The results of the experiments show that the average PSNR and SSIM of this method on the outdoor synthetic dataset are improved by 1.256 7 and O.O30 8，respectively，compared with the optimal results of the current mainstream methods.Meanwhile，in the test on the real-worlddataset，the PIQE，RI，and VIreach the optimal indexes of 21.471，0.9711 and 0.90o 3.EDFDNet achieves good results in both realism enhancement and visibility restoration，and is suitable for dehazing real-world non-uniform haze images.

Key words： real-world haze； non-homogeneous dehazing； feature fusion； edge feature enhancement;generative adversarial network

1引言

在自然场景中，雾霾中的颗粒物和微小水滴会对光线产生散射和吸收作用，导致图像的色彩和对比度显著降低，细节和边缘变得模糊，场景中的物体无法准确再现，影响了图像分割1、物体检测和识别[2等高级视觉任务的有效性。（剩余18626字）