Target-Dependent Scalable Image Compression Using a Reconfigurable Recurrent Neural Network

Conventional human-centric image compression techniques are optimized for human visual perception, and generally evaluated by metrics such as MSSSIM PSNR. On the other hand, task-centric that target deep neural networks (DNN) inference focus on understanding images, measured accuracy. As images should be encoded decoded differently depending metric, existing learning based have focused designing an independent network is either one of two targets. However, these require metric to determined during a training phase, separate DNN models trained different applications, allowing no scalability between human- compression. In this paper, we propose target-dependent scalable pipeline, where compressed can in real time interpolated perceptual quality This dynamic supported reconfigurable recurrent (RNN) dynamically change its flow according interpolation parameter. The experimental results show proposed method achieves comparable classification accuracy individually targets, respectively, while providing targets