Geometry Based Image Segmentation

Geometric data analysis extracts edges, lines, curves, and shapes to provide a feature-based representation of an image. An active learning framework is then used to learn to differentiate between objects of interest and background clutter. Once the learning algorithms are sufficiently trained, machine learning tools can be applied in parallel to provide high-throughput processing of large datasets

Geometry Based Feature Extraction

Images and other complex datasets can be represented as feature sets based on features such as edges, lines shape, structure, and other topological features of data. In this approach, the image is transformed into geometric representations based on key features that exist across different scales. This representation is often well suited for noisy datasets where conventional image analysis approaches tend to perform poorly. The following image shows the evolution of the segmentation process.

geomsegmentation1

Active Learning Approach to Automatic Segmentation

Ensuring accurate segmentation is critical in many image analysis applications. GDA ensures accurate and reliable results through an active learning framework that allows an expert user to label a small subset of the feature space and then adaptively train the system to properly differentiate objects of interest from the background clutter. This interactive process allows the user to refine the training algorithm and high accuracy with confidence with minimal effort. Using the active learning  interface shown below, expert users can label a small number of example segments to train the system to correctly label the whole image.

 

Active Learning framework for image segementation

Cloud-based Scalable Processing

Developed to run as a cloud-native application, this image segmentation solution was designed from the ground to  support arbitrarily large datasets. Once the active learning process is sufficiently trained, multiple images can be analyzed in parallel through a scalable, cloud-based processing framework. Built as a container-based, cloud native application, this solution leverages modern python tools to provide a scalable, rapid and accurate solution for image segmentation problems. A centralized data management and reporting system aggregates the results and makes the segmented images available for further processing as part of an extended analysis pipeline.

RootShape Engineering Diagram

Bringing It All Together

Once the active learning process is sufficiently trained, multiple images can be analyzed in parallel through a scalable, cloud-based processing framework. Built as a container-based, cloud native application, this solution leverages modern python tools to provide a scalable, rapid and accurate solution for image segmentation problems.

Plant root depth analysis based on geometric image segmentation.

This solution leverages GDA’s capabilities in data analysis, machine learning and scalable computing for high through segmentation of datasets with noisy and cluttered images. Combining these capabilities allows us to accurately and rapidly segment and label large image datasets. Since segmentation is often the first step of a larger image analysis framework, this solution is implemented using an API framework to allow this capability to be easily integrated into these existing processing pipelines and workflows.

Relevant Projects

Geometry based image segmentation is used for image analysis in these projects at GDA.

RootShape: Automated Analysis of In Situ Fine Root Images (DoE Phase II SBIR, 2021)

RootShape provides automated segmentation and annotation utilizing differential geometry applied to digital rhizotron and minirhizotron images to recognize and filter root-like regions from background. Then a low-dimensional vectorization of these segmented regions is derived, which is used to train a very simple classifier with limited user intervention.