Conclusions and Future Work
The computational efficiency of calculation of image descriptors increases with increasing computation and it increases with decreasing transmitted data. The implemented approach allows the processing algorithms to be isolated from the specific elements of computing environment which provides a certain flexibility for software implementation and testing.
We plan to develop a computing cluster on the basis of a tuberculosis dispensary and then to build a dedicated computational GRID infrastructure for supporting large-scale image computing.
Because of computational capacity of image processing algorithms and very large amount of visual data, it makes sense to use an approach capitalizing on a distributed computing architecture. A number of tasks were stood out including data communication, efficient allocation of computing resources, synchronization and optimization of computing process.
We tested the suggested software in a local network (100 Mbit/s) with 15 computing nodes. The cluster node configuration included computers with an 1.6GHz CPU, 2Gb RAM, Windows XP Professional. The Firebird was used as a database server. Experimentation revealed that the critical bottleneck in the environment was the data communication because of the large volume of transmitted visual data. Each compute node should receive a chunk of data, to compute them and send the result to the database. For minimization of data transfers, a special computational procedure was developed and implemented. In order to secure data safety and provide an authorized access to the private medical information we plan to utilize the security mechanism of the UNICORE GRID middleware.
An experimental study of indexing software on the test image sets containing 1000, 2000, 3000 and 4000 images has demonstrated interesting results. For the indexing of small number of medical images (up to 4000), it is best to use a dedicated computer without the organization of the distributed computing environment. In the GRID infrastructure, time is also spend on the exchange of information between the service components. A substantial contribution is the time for interaction with the database, which is located on a dedicated computer. The processes of the reception and transmission of the images or their descriptors also have an effect on the overall computational time too .
But with increasing medical images databases, using distributed computing becomes viable, and the benefit from the proposed scheme of calculation is becoming more noticeable.
|URL for further information||http://safonov.by/grid/index.html|
|Keywords||Lung image, Content-based Image Retrieval, Indexing, Distributed computing|