Meta Transfer Learning For Zero Shot Super Resolution

We found that simply employing transfer learning or fine-tuning from a pre-trained network does not yield plausible results.

Meta transfer learning for zero shot super resolution. To address both issues zero-shot super-resolution ZSSR has been proposed for flexible internal learning. Meta-Transfer Learning for Zero-Shot Super-Resolution CVPR 2020 - JWSohMZSR. CVPR 2020 论文Meta-Transfer Learning for Zero-Shot Super-Resolution.

74 Jae Woong Soh Sunwoo Cho Nam Ik Cho. 512 and mini-batch size to 16. Next we start meta-learning for these SR networks in accordance with iterative steps in Algorithm.

Still to quickly optimize the network MZSR was limited to use a. 75 Yong Guo Jian Chen Jingdong Wang Qi Chen Jiezhang Cao Zeshuai Deng Yanwu Xu Mingkui Tan. Despite their remarkable performance based on the external dataset they cannot exploit.

However they require thousands of gradient updates ie long inference time. Meta-Transfer Learning for Zero-Shot Super-Resolution. Present Meta-Transfer Learning for Zero-Shot Super-Resolution MZSR which is kernel-agnostic.

SR which exploits the power of Deep Learning but does not rely on prior training. They added a meta-transfer learning phase to exploit the information of the external dataset which decreased the number of the steps required at runtime. Precisely it is based on finding a generic initial parameter that is suitable for.

For meta-learning we still use DIV2K dataset and use 5 inner gradient update steps for 2 line 7 in Algorithm 1. CVPR 2020 Jae Woong Soh Sunwoo Cho Nam Ik Cho. Precisely it is based on finding a generic initial parameter that is suitable for internal learning.