Image Feature Extraction And Index Construction Based On Dataflow

In order to be able to find interesting target image in the existing flood of pictures, we need to establish an efficient index structure depending on the image features. However, how to use modern multicore processor platforms massively parallel computing power to quickly build the index structure exists a difficult challenge. Thus, data flow programming model as a field-oriented programming model is proposed to simplify the programming of multicore architecture, enhances application portability and scalability, however, there are some flows in usability.For the problem of availability about data flow programming, I designed and implemented a new hybrid programming model, through the experiments of the hot application image indexing to prove the availability of the model. The system extends the original COStream dataflow programming language, increases an interface and realizes the hybrid programming model between COStream and C++. In view of the image indexing, first analyze the image of Scale Invariant Features Transform feature value extraction, which is suitable for the flow of data parallel computing part COStream technique was used to realize, the other part keep the original C++ program unchanged, to combine the two pats, can generate efficient parallel execution of the program. Then we analysis the image index build process based on residual quantitative, in the same way, which is suitable for the part of the data flow parallel COStream technique was used to realize, the rest part remain the C++ program, finally generate efficient executable program.I choose common X86-64 as the experiment platform to perform the experiment and analysis programs performance. The experimental results show the feasibility of this hybrid programming model.