AI-based measurements For Imaging

Summary

NIST streamlines deriving measurements from terabyte-sized images and improves image-based measurement accuracy via web-based applications addressing image analyses and artificial intelligence (AI) based modeling.

Description

The problem addressed in this work lies in enabling scientists to apply artificial intelligence (AI) based models to deriving measurements from terabyte-sized images. This problem is important as image-based measurements can become more accurate by introducing supervised AI-based models instead of using the traditional machine learning (ML) based models.

The challenges of introducing supervised AI-based models to domain scientists and assisting them with AI-based modeling are in:

1. the complexity of mastering constructions of AI-based models, training and evaluation of AI-based models, inference using the built models, and leveraging hardware acceleration,
2. the variety of many image pre- and post-processing steps in order to train AI-based model and use them for inferencing, and
3. the time period it takes to deliver cutting-edge, research-prototype, AI-based models to the hands of measurement scientists.

Major Accomplishments

We research the integration of the NVIDIA DIGITS framework with the web image processing pipeline (WIPP) system developed by NIST.

We engage in the development of AI model representation (i.e., Open Neural Network Exchange or ONNX format) in order to facilitate the inter-operability of AI-based services and their definitions on top of AI platforms, such as TensorFlow, Caffe2, Microsoft Cognitive Toolkit (CNTK), PyTorch, or Apache MXNet.

Associated Product(s)

Web Image Processing Pipeline (WIPP) software: https://isg.nist.gov/deepzoomweb/software/wipp

NVIDIA DIGITS: https://developer.nvidia.com/digits

Publications:

https://isg.nist.gov/deepzoomweb/resources/csmet/pages/cnn_bcars_rules/cnn_bcars_rules.html