Convolutional machine learning models using multiple tiling mechanisms

Abstract

As described herein, various embodiments of the present invention disclose techniques that improve efficiency of performing image-based machine learning operations on large images while limiting accuracy drawbacks of partial processing of those large images by using composite-tiled image embeddings for composite-tiled images generated by merging multiple tiled images that are generated using multiple tiling mechanisms. For example, in some embodiments, given an input image that comprises R image regions, each tiled image comprises N selected image regions of the R image regions that are selected in accordance with a tiling mechanism (where N<R). In this way, given T tiling mechansisms, T tiled images are generated, and the T tiled images are merged to generate a composite-tiled image. Accordingly, by using T tiling mechansisms, various embodiments enable reducing the size of feature data provided to an image processing machine learning model by selecting non-holistic subsets of an input image.