一种基于生成对抗网络的无人机图像去雾算法
UAV Image Dehazing Algorithm Based on Generative Adversarial Network

无人机所采集的图像容易受到雾霾、雾气等阴霾天气干扰，造成图像质量下降。针对阴霾天气下无人机采集图像的质量下降问题，提出了一种新颖的基于生成对抗网络的图像去雾方法。本方法设计了新式生成网络和判别网络，生成网络由多层编码器和解码器对称分布构成，判别网络由全卷积网络构成，为了提高生成图像的清晰度，引入了一种新的对抗和平滑损失函数来优化整个网络。最后，通过大量实验表明，基于本文方法进行图像去雾取得了良好的效果，在结构相似度和峰值信噪比等评价指标以及主观视觉效果上优于已有的图像去雾方法。
The image collected by UAV is easy to be disturbed by fog, which leads to the degradation of image quality. Aiming at the image degradation of UAV in foggy scenes, a novel image defogging method based on generative adversarial network is proposed. Anew generator and discriminator are designed. The generating network consists of multi-layer encoder and decoder; then the discriminator network consists of fully convolutional network. In order to improve the clarity of the generated image, a new loss function is introduced to optimize the whole network, including adversarial loss and smooth loss. Through training and testing, it can be concluded that the image defogging method based on generative adversarial networks has achieved good results, and it is better than traditional methods in structural similarity index measurement (SSIM) and peak signal to noise ratio (PSNR).