Abstract
Identifying tiny objects with extremely low resolution is generally considered a very challenging task even for human vision, due to limited information presented inside the object areas. There have been very limited attempts in recent years to deal with low-resolution recognition. The existing solutions rely on either generating super-resolution images or learning multi-scale features. However, their performance improvement becomes very limited, especially when the resolution becomes very low. In this paper, we propose a Representation Learning Generative Adversarial Network ( RL -GAN) to generate super image representation that is optimized for recognition. Our solution deals with the classical vision task of object recognition in the distance. We evaluate our idea on the challenging task of low-resolution object recognition. Comparison of experimental results conducted on public and our newly created WIDER-SHIP datasets demonstrate the effectiveness of our RL -GAN, which improves the classification results significantly, with 10–15% gain on average, compared with benchmark solutions.
Highlights
Recent advances in object recognition are largely stimulated by deep learning techniques, such as ResNet [1], DenseNet [2] and SeNet [3], which learn deep representations from regions of interest (RoIs) and perform classification
In our work, aiming at achieving high classification accuracy directly from LR images, we propose a Representation Learning Generative Adversarial Network (RL-generative adversarial network (GAN))
The main contributions of this work are: 1) We propose a RL-GAN architecture to enhance the discriminability of the LR image representation resulting in comparable classification performance with that conducted on HR images
Summary
Recent advances in object recognition are largely stimulated by deep learning techniques, such as ResNet [1], DenseNet [2] and SeNet [3], which learn deep representations from regions of interest (RoIs) and perform classification. In our work, aiming at achieving high classification accuracy directly from LR images, we propose a Representation Learning Generative Adversarial Network (RL-GAN). Bai et al [24] proposed a super-resolution RoIs based generative adversarial network, which consisted of two modules, i.e., the generator, which was a super-resolution network to up-sample LR images into HR ones and recover the detailed information for more accurate detection, and the discriminator, which was a multi-task network for classification and bounding box regression. The transfer learning based methods transfer external knowledge in high-resolution images to improve the performance for low-resolution object recognition. Bulat and Tzimiropoulos [45] proposed a multi-task deep model to simultaneously learn face super-resolution and facial landmark localization trained using a generative adversarial network (GAN).
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
