TilePredictor: Image and Pixelwise Classification with Tile-based Convolutional Neural Networks (CNNs)

About this technology

Overview

TilePredictor is a python library designed to rapidly construct and evaluate CNN-based image classifiers for image and pixelwise classification problems. In particular, TilePredictor is designed to generate pixelwise predictions using image-based CNN classifiers which generate a single prediction per image for large images such as those typically encountered in remote sensing and astronomy. TilePredictor generates pixelwise predictions by efficiently spatially aggregating predictions generated by convolving an image-based CNN over tiles extracted from a target image. Additionally, TilePredictor wraps essential CNN training and validation functions to enable rapid construction, evaluation and deployment of both pre-trained and newly-instantiated image-based CNN models to new image data with minimal effort.