Abstract: In recent years, the application of deep convolutional neural networks in UAV aerial image super-resolution tasks has made a huge leap in the performance of UAV aerial image super-resolution. However,super-resolution methods based-on convolutional neural network rely on specific training datasets, which are typically constructed by downsampling images using a fixed bicubic kernel. When the processed image does not meet this “ideal” situation, its performance will drop dramatically. This paper proposes a self-supervised UAV aerial image super-resolution reconstruction method based on non-local mining reconstruction network (NLMRN) to solve this problem. NLMRN does not require external datasets for pre-training, only an input image. It exploits the non-local reproducibility of the internal information of UAV aerial images by downsampling the input image itself to obtain a lower resolution image for training. To better learn non-local repeated features, we use the non-local context mining block (NLCM) to establish relationships between non-local regions and select a subset of global feature maps to supplement each specific location to obtain precise details and texture reconstruction. NLCM effectively makes up for the shortcoming that the convolution operation can only process one local neighborhood at a time. Through extensive experimental verification, NLMRN is significantly better than other advanced super-resolution methods when processing UAV aerial images under “non-ideal” conditions.

Key words: UAV aerial, super-resolution, non-local reproducibility, global feature map