High-Resolution Image Inpainting through Multiple Deep Networks

For the operation and aerial photography of the UAV, it is important to identify the blindspots and observe the details on the ground. But limited by the camera resolution, small or fuzzy objects can not be effectively observed. Therefore, repairment of high-definition images has become one of the important problems to be solved. In recent years, the development of the deep learning method has effectively solved the loss and blurring of images, but because of the difficulties in training and the speed of calculation it can only be used with low-pixel images. Therefore, we propose a method for superimposing images first with the content and textual recovery for the defaced area. We use unsupervised learning GANs and trained VGG network to restore holes and missing areas of the image, and then enlarge it through CNN method. Our preliminary results show that high resolution image restoration speed has been greatly improved, and details become sharper than using traditional method.