Skip Header NavigationIntranet 
CENTER FOR EMBEDDED NETWORKED SENSINGContactDirectionsEmploymentEventsNews
HomeAbout UsResearchEducationResourcesPeople

Research Project

Visual detection and classification of structures of interest in minirhizotron images

Technology > Systems Area Projects > Visual detection and classification of structures of interest in minirhizotron images

On this page: Overview | Approach | Systems/Experiments | Accomplishments | Future Directions | People

Overview

This research focuses on developing algorithms to be used for manual, semi-automatic, and possibly fully automatic analysis of minirhizotron images to detect structures of interest from suitably labeled examples. We aim to develop a suite of algorithms to support of the existing classification modality and to extend toward semi and fully automatic classification.

Approach

The goal of this project is to produce software tools for the analysis of images obtained from a minirhizotron. The system consists of a suite of C/C++ algorithms including image processing algorithms, learning-based classification, and fully automatic image analysis algorithms.

Systems/Experiments

We have created a cross-platform image analysis software with a modular filter plug-in architecture (Figure 1). This software currently includes filters for scale space ridge detection [Lindeberg], canny edge detection, anisotropic diffusion, and a difference of gaussians filter.

Figure 1: The minirhizotron software interface uses OpenCV and QT to provide a modular plug-in architecture and cross-platform compatibility.

Figure 1: The minirhizotron software interface uses OpenCV and QT to provide a modular plug-in architecture and cross-platform compatibility.

Evaluating these filters on a variety of minirhizotron images led to the choice of ridges as the basic feature for classification. At each ridge we attached a simple descriptor consisting of the ridge saliency, local color information, average orientation along the ridge, and length.

Approximately 60 images were manually classified by masking out regions of the images which did not contain structures we were interested in. This resulted in 411 ridge segments which were of interest. An equal number of negative examples were selected at random from dataset. These were then split into a training set consisting of 320 positive and 320 negative examples, and a test set of 91 positive and 91 negative examples. A support vector machine with a radial basis function kernel was then evaluated on this data. The results may be found in Figure 3.

Accomplishments

Figure 2: The top 150 bright and dark roots, detected at all scales.

Figure 2: The top 150 bright and dark roots, detected at all scales. The interesting structure travels from left to right in the upper portion of this image, and is detected as long continuous ridge segments. Different colors correspond to different ridges.

Our scale space ridge detection implementation allows us to find and rank ridge structures in general images. In Figure 2, it may be seen that ridges corresponding to the root were found within the top 150 ridges.

Additionally, we have trained a support vector machine and achieved the results shown in Figure 3. These results suggest that it may be feasible to classify structures of interest in minirhizotron images using a good feature descriptor attached to a ridge. Figure 3: ROC curve for an RBF SVM trained as described above.

Figure 3: ROC curve for an RBF SVM trained as described above.

Future Directions

In the future, we first plan on exploring various feature descriptors attached to the ridges. Some potential additions include mean curvature, local texture information, and the standard deviation of scale along the ridge. We will also improve our manual classification tools in order to make it easier to collect a large dataset of classified images. It is likely that it will be necessary to introduce a higher level vision model which performs cue integration along ridges exploiting knowledge learned about the structures from many examples. These steps will move us closer to fully automatic classification of interesting structures in minirhizotron images.

People