1. Short overview
A new nonparametric Harris-affine detector is introduced here. This is an image processing algorithm for extracting a particular kind of image features. The new proposed implementation automatically tries to select best features with respect to local-to-global image properties in a scale-space domain. An unusual parallel GRID implementation has been developed to avoid unbalanced computational workload distribution among different processors.
Provide a set of generic keywords that define your contribution (e.g. Data Management, Workflows, High Energy Physics)
parallel image analysis, MPI feature detector, scale-space theory, no-parameters algorithms, Genius
4. Conclusions / Future plans
Good results have been obtained considering that for some sections the parallelism degree is bounded by the numbers of used scales and also the bottleneck of very heterogeneous data. From a technical point of view, our application needs an useful installation of FFT library; such installation has been inquired to the PI2S2- Grid technical team, and it will be running by the next few days. We will discuss the resource required for it, the performance and its scalability on GRID paradigm.
The proposed algorithm consists of the following steps: image enhancement and feature mask computation by using z-scored local windows, simple Harris-corner extraction and selection, and refinement of the final result by an iterative procedure computed on every feature without computational approximation. The algorithm uses statistical filters with a variety of kernel, which cause a bottleneck on a serial implementation. Right now, our application has been developed under MPI paradigm and a corresponding porting for PI2S2-GRID is under construction and we foresee that the final Grid version will be tested by a couple of weeks. The system will use the support of Genius for a dissemination on a naïve scientific community and also to display the results of large data. Given the latency of standard network we assume a improving of the performance with the use of Infiniband network. The efficiency of our MPI methodology has been test on a set of images and it has been evaluated about 80%.