Image Segmentation: Delineation of Rock Fragments Using Compressed Random Features

This dissertation presents several practical 2D image segmentation algorithms
for delineation of rock fragments in images of rock piles. Such analysis is important for improving the efficiency and safety of underground mines. Recent developments in 3D imaging systems have resulted in significant improvements in fragmentation analysis. However, many mines have already installed 2D imaging systems and 3D imaging is not a feasible alternative for them. The disadvantage of using 2D images of rock piles is that input from a human operator is often required.

The dissertation proposes the use of compressed random features for image
segmentation. These features are derived by dividing images into square patches
and then compressing the patches using Haar-like features. To determine whether these features are suitable for image segmentation a frequently applied algorithm, Normalized Cuts, is evaluated using the compressed features and features derived from a filter bank. The compressed features result in better segmentations, with two to three times faster processing of features.

Normalized Cuts has some serious disadvantages for rock fragmentation
analysis. An alternative approach is a support vector machine (SVM) classifier,
which determines whether each image patch contains a rock fragment boundary.
The optimum values for the patch size and number of compressed features are
determined by evaluating the trade-offs involved in using an SVM. For the rock
images in this thesis, the optimum patch size is 15×15 and 20 compressed features is sufficient. Compared to uncompressed image patches, the compressed features give a 12.75 times compression and 25 times improvement in processing times.

Another method explored is using cluster centres learned for each class of
pixels present in an image. K-means is used to cluster the proposed compressed
features in training images and then nearest neighbour classification is used on
features from test images; this gives a simple and quick training procedure and
very fast classification times. For rock segmentation, 16 cluster centres are opti-
mum. However, the best accuracies observed for this approach (78%) are not as
good as those using an SVM (85%).

The dissertation describes the development of a robust nearest subspace classification algorithm. This algorithm is evaluated for tasks including identification of hand written digits and face recognition. Robust subspace learning is then applied to classification of image patches for image segmentation and rock fragment delineation. While the results are very good for hand written digits and face recognition, they are poor for image segmentation and rock delineation.

Two dictionary learning algorithms are proposed and evaluated. The first
seeks a low rank dictionary and proposes a novel regularization term. The second uses a Sparse Group Lasso-like approach to match the coefficients’ structure to the data’s class structure. These algorithms give promising results both for segmenting images from the BSDS500 dataset and for rock delineation.

The results presented suggest several promising avenues for future research.
These include exploring the use multiple images taken with different lighting
to reduce delineation errors; and improvement of the proposed nearest cluster
classifier. A further possibility is investigation of linking partial boundaries when
using compressed random features.