Parallel Computing for Medicinal Plant Identification System Using Fuzzy Local Binary Pattern

Abstract

As biological image databases are increasing rapidly, automated species identification based on digital data is of great interest for accelerating biodiversity assessment, research and monitoring. This research applied high performance computing (HPC) on medicinal plant identification system. We propose parallel computing on medicinal plant image processing using Fuzzy Local Binary Pattern (FLBP). The main goal of the research was to measure the efficiency of parallel computing on medicinal plant image processing and evaluation whether this approach is reasonable for handling large data sets. This research proposes two models of parallel design to identify medicinal plant. The first model used data parallel design and the second model used task parallel design on FLBP process. Both of model are applied on the computer cluster, which consists of eight computers with the same spesicification. The development of the parallel design used the message-passing model with MPI library and the C/C++ language programming. The parallel computation performance was evaluated by speed up, efficiency and iso-efficiency. The experimental result shows that both of the parallel design models can reduce the computing time of the image feature extraction on a medicinal plant identification system. The values of the speedup on the first model by the time of the extraction was 1 440 image data by using 8 processors, is 7.64 with efficiency value of 0.95. The second model uses 8 processors generated the value of the speedup is 6.9 with efficiency value 0.87. The result of the comparation values between the speedup and the efficiency of both design model shows that first model has better performance by the time of the extraction 1 440 leaf images. This is affected by the process of image area dividing on the second model that requires more complex of communication cost. We analyzed that in the second model if we add processor would effect on communication cost. This condition did not occurred on the first model, so that the speedup and efficiency are better than the second model. The experimental result of query image extraction shows that the speedup on the first model is 6.73, with efficiency value of 0.84. The second model produces speedup value is 7.96, with efficiency value of 0.99. The different between speedup value and efficiency produced is affected by the dividing of data from those two parallel models. The first model divides 20 combination of the operator and FLBP threshold. The process of data dividing on first model enables unideal data partitioning. The ideal data dividing is at the time of each processors compute the data with the same volume. The unideal data dividing may cause an idle condition. The idle condition can affect the performance and the parallel time generated by the first model. The dividing of the data on second model by dividing the image area equal to the processors used. The data dividing on the second model is ideal condition, so that the speedup and efficiency are better than first model.