Scale-invariant image editing

Abstract

We introduce a system architecture for image editing which decouples image filtering from the image size, resulting in a system which allows interactive editing with constant response times, independent of image sizes. Scale invariance means filters are designed to allow scaled rendering from a pre-computed image pyramid, approximating the result of the filter when rendered at full resolution and scaled afterwards. Our implementation of the proposed architecture allows interactive editing on all image sizes with minimal hardware requirements, the least powerful device tested with a 300 megapixel image was based on a dual core ARM Cortex A9 clocked at 1Ghz. The architecture is based on a graph based image editing approach, extended by scaled rendering for all filters. The filter graph is exploited to allow automatic configuration of filter properties and conversion between color spaces, which simplifies filter implementation and increases performance. The handling of image data is based on tiles and a tile cache allows to manage memory requirements and increase interactive performance. The implementation is provided as a portable library written in c and can provides interactive editing on device as slow as last generation smartphones, while at the same time exploiting the performance available to current multi core processors, using effective multithreading. In this work we explore both the architectural details that make this possible as well as the properties of common image editing filters, regarding the required scale invariance. We also examine possible approaches that can be followed to implement practical filters for such as system. Finally, the implemented architecture and filters are extensively tested for performance and accuracy and the results are examined.