Acceleration Of X-ray Computed Tomography Reconstruction For Rice Tiller

The number of rice tillers was an important phenotype parameter, which determined panicle number, a key component of grain yield. On the Multiple Rice Phenotype Parameters Extraction System, a fan beam X-CT (X-ray Compute Tomography) system was used to scan transverse section of rice tillers. The filtered back-projection algorithm was used to reconstruct the transverse section image of them. However, the reconstruction was time-consumed. In order to speed up the measurement and real-time reconstruct and display the image, two approaches were adopted to accelerate image reconstruction: ROI (Region of Interest) identification and GPU (Graphic Process Unit) parallel computing. The ROI identification algorithm was to previously identify the region of interest in the reconstructed image.In this way, the amount of the reconstructed pixels was reduced dramatically which could directly speedup the image reconstruction. Besides, the filtered region of interest in Sinogram could be restricted after identification of ROI in the reconstructed image. This thesis also adopted the CUDA (Compute Unified Device Architecture) programming to implement the GPU parallel computing for image reconstruction.The results showed that the ROI identification algorithm could speedup 4~10 times than FOV reconstruction (15 rice samples in mature stage). It dramatically accelerated the back-projection stage, while the pre-weighted convolution filtering could be slightly accelerated. The GPU-based image reconstruction was approximately 180 times faster than the CPU-based one. The pre-weighted convolution filtering and back-projection could be accelerated approximately 50 times and 250 times, respectively. Moreover, the different between the GPU-based reconstruction image and CPU-based reconstruction image was not significant. For the reconstructing image with 768×768 size, the consumption time was about 100 ms. Combing the ROI identification algorithm and the GPU parallel computing, the results showed that the image reconstruction could be accelerated approximately 450 times. The pre-weighted convolution filtering and back-projection could be accelerated approximately 80 times and 2000 times, respectively.