Name
Title | ||
|---|---|---|
Haralick's Texture Features Computation Accelerated by GPUs for Biological Applications. |
Abstract | ||
|---|---|---|
x k P f k f x y x k P k P f Abstract This poster presents the speedup of the computation of co-occurrence matrices (co-matrices) and Haralick Texture Features (features), as used for analyzing microscope images of biological cells, by general-purpose graphics processing units (GPUs). In a pipeline of automated image analysis algorithms the feature computation is the most costly computing part. The computing time of the algorithm without acceleration amounts to several months. Hence, a massive speedup is required. Analyzing the features results in a graph showing the dependency of the feature computations on intermediate results and on other features. With the dependency graph the optimal order of the feature computation could be determined which saved costly double computations. Analysation of co-matrices showed that they are sparsely filled, and for a highly parallel approach they consume too much memory. We reduced the size of a full co-matrix by removing all rows and columns filled with zeros. This reduction strategy allowed us to keep up to two hundred co-matrices in the memory of an ordinary graphics card with direct memory access. For each single cell image 20 co-matrices with different orientations are generated. Altogether, the features of 8 cells can be computed in parallel, requiring the calculation of 160 co-matrices. To reduce the complexity of the feature computation 24 kernel functions are used on the GPU and each one maps all co-matrices to the parallel computing architecture of the GPUs. On a single node of a cluster, a speedup of 500 was obtained compared to an unoptimized software version, and a speedup of 46 was obtained compared to an optimized software version. The GPU implementation reduces the computational time from half a year to around 9 hours, which opens up a new research application in the field of biological image analysis. |
Year | DOI | Venue |
|---|---|---|
2009 | 10.1007/978-3-642-25707-0_11 | HPSC |
Field | DocType | Citations |
Graphics,Computer vision,Co-occurrence matrix,CUDA,Computer science,Parallel computing,Thread (computing),General-purpose computing on graphics processing units,Artificial intelligence,Stream processing,Speedup,Computation | Conference | 1 |
PageRank | References | Authors |
0.37 | 3 | 7 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Markus Gipp | 1 | 23 | 1.89 |
| Guillermo Marcus Martinez | 2 | 1 | 0.71 |
| Nathalie Harder | 3 | 114 | 17.57 |
| Apichat Suratanee | 4 | 9 | 2.49 |
| Karl Rohr | 5 | 33 | 7.02 |
| Rainer König | 6 | 53 | 3.84 |
| Reinhard Männer | 7 | 1 | 0.37 |