| Summary: | 碩士 === 長庚大學 === 資訊工程學系 === 105 === With more and more Drosophila Brain images (Driver and Neuron) generated, it is an important work to understand the Driver-Driver similarity, Neuron-Neuron similarity and Driver composition of Neurons.
For the similarity between the Driver images, the similarity between the neurons images and the similarity between the Driver images and the neurons images, we summed up the significance of the three similarities: 1. Using the similarity between the neurons images, finding a block that deliver messages to each other. 2. Using the similarity between the Driver images and the neurons images, we can see which Driver’s genes performance is related to neuron cells, and we can also use genes to control the performance of neuron cells. 3. Although the importance of the similarity between the Driver images is less important than the similarity between the Driver images and the neuron images, and the similarity between the neuron images, but the function is that when the biologist is doing the experiment, choosing best Driver to do the experiment.
All of more than 10,000 Drosophila Driver images and 28,000 Drosophila Neuron images have been compressed from original image files into .CRS files by using CRS scheme. CRS is a traditional matrix compression technique that removes nonzero values and removes newline symbols, so the read speed is quite fast. In this paper, we have two comparison methods: complete comparison and fuzzy comparison. The complete comparison is when original coordinates is completed matched then the VALUE will be record, and fuzzy comparison will be based on the original coordinates to give it a range of an error value, such as positive and negative.
Then, these compressed images were compared with each other by using the image-matched method. More than 400 million pairs should be compared for N2N; more than 50 million pairs for D2D; more than 250 million pairs for D2N.
In recent years, the processing power of Graphic Processing Units has been a breakthrough, and has been applied to many research areas, how to use the graphics processor technology to improve the image comparison performance, will be a very important research topic. Drosophila image comparison is the longest time spent in the whole architecture, not just the number of comparison times, each index of a single neuron and a single driver images can be as high as 60 million and 40 million. So a large number of Drosophila brain neurons and drive images used to do comparison is the need to consume a lot of time, so in this paper, we propose a set of GPU algorithms to reduce the consumption of image comparison time.
We use the machine's CPU model for the Intel Xeon E5-2650 v2. In the GPU part, we use the NVIDIA Tesla K20m and NVIDIA Tesla K40m, and in the driver to driver comparison, K20m can be speed up to 30 times faster, and K40m can be up to 100 times; in the driver to the neuron comparison, K20m fastest 10 times, K40m the fastest 30 times; the last in the neuron to neuron comparison, K20m fastest 30 times, K40m can be up to 66 times.
|
|---|
