Deep Learning Methods for Streaming Image Reconstruction in Fixed-camera Settings
Johan Förberg, presenterar sitt examensarbete
A streaming video reconstruction system is described and implemented as a convolutional neural network. The system performs combined 2x super-resolution and H.264 artefacts removal with a processing speed of about 6 frames per second at 1920×1080 output resolution on current workstation-grade hardware. In 4x super-resolution mode, the system can output 3840 × 2160 video at a similar rate. The base system provides quality improvements of 0.010–0.025 SSIM over tradtional Lanczos filtering. Scene-specific training, in which the system automatically adapts to the current scene viewed by the camera, is shown to achieve up to 0.030 SSIM additional improvement in some scenarios. It is further shown that scene-specific training can provide some improvement even when reconstructing an unfamiliar scene, as long as the camera and capture settings remain the same.