Single-Image Dehazing on Aeria Imagery Using Convolutional Neural Networks

Kurzfassung

Aerial images are widely used in various applications such as land-use planning, environmental studies, sustainable development projects, smart cities and traffic management systems. The atmospheric conditions such as hazy weather at the flight time can affect the resolution and quality of the aerial image and consequently influence the performance of the aforementioned tasks. Haze contains floating particles in the air which can result in image quality degradation and visibility reduction in airborne data. Haze removal Task has several applications in image enhancement and can improve the performance of automatic image analysis systems, namely object detection, and segmentation. In this thesis, we propose a well-performing method to dehaze aerial images using Convolutional Neural Network (CNN). Despite rich haze removal literature in the ground-level imagery, there is a lack of methods specifically designed for aerial imagery. Considering the fact that there is a characteristic difference between the aerial imagery domain and ground one, we may not obtain the best dehzing results by applying a ground-level imagery dehazing method with no appropriate modifications on aerial images, so, a domain adaption is needed which is proved by our experiments. Investigating through different state-of-the-art ground imagery dehazing methods, All-in-One Dehazing Network (AOD-Net) is chosen as the baseline for adaption. Its outperforming results in ground-level imagery as well as its straightforward structure, flexible for further adaptation, motivate us to utilize this network in our experiments. The domain transfer is done by training the network on aerial imagery. To do so, an aerial hazy image dataset is needed. To the best of our knowledge, there is no publicly available hazy aerial image dataset available and therefore, we create a new synthetically-hazed aerial image dataset in both homogeneous and inhomogeneous versions using different assumptions and computational strategies. For the homogeneous case, we assume to have the same ground height for all image pixels, while in the inhomogeneous case, a random Digital Elevation Model (DEM) is created and used in the hazy image generation procedure. After training the network on the generated dataset, we test our model on natural as well as the synthetically-hazed aerial images. Both qualitative and quantitative results of the adapted network show an improvement in dehazing results. We show that the adapted on our aerial image test set increases Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSim) by 2.2 dB and 9%, respectively. Despite the improvement, the color suppression problem is still seen in the adapted network. This effect can also be realized from the Kullback-Leibler Divergence (KLD) of the color histograms between dehazed and ground truth images. Hence, we further modify the network by making it keep the Mean Squared Error (MSE) between the dehazed image and the ground truth one as minimum as possible, not only in the time domain but also in the frequency domain. The results show an improvement of the dehazed results by 2.24 dB in PSNR and 16.6% in SSim.