Small Object Detection Networks Based on Classification-Oriented Super-Resolution GAN for UAV Aerial Imagery

Small object detection is challenging for lack of discriminative information compared to medium, large objects. In this paper, we proposed small object detection networks based on classification-oriented super-resolution generative adversarial networks (CSRGAN) and validate them on public UAV aerial imagery benchmark. Through appending classification branch and introducing classification loss to typical SRGAN, generator of CSRGAN is trained to reconstruct realistic super-resolved (SR) images with classification-oriented discriminative features from low resolution images while discriminator is trained to predict true categories and distinguish generated SR images from original ones. Besides, VGG19 based feature-level content loss is applied to recover clearer and sharper contours in SR images, which is critical to object classification. Experiment results prove the classification-oriented enhancing effect of CSRGAN and the positive function of VGG19 based feature-level content loss.