CUDA based Iris Detection based on Hough Transform
2 likes1,238 views
The document details a CUDA-based iris detection project utilizing the Hough transform for real-time segmentation of the iris and pupil. It includes performance measurements comparing CPU and CUDA implementations across various image sizes and highlights the efficiency gains achieved through shared memory and optimization techniques. Future improvements are suggested to enhance processing speed and eliminate CPU-based pre-processing.
CUDA based Iris Detection based on Hough Transform
- 1. CUDA based Iris Detection based on Hough Transform G P - G P U C o u r s e P r o j e c t J u s t i n a s M i š e i k i s
- 2. Goal Eye detection in a color or grayscale image Detect boundaries of iris and pupil for the segmentation Make the detection real-time
- 3. CUDA parts Hough transform for circles Peak detection Image adjustment Gamma correction CPU parts Canny edge detection - OpenCV Histogram calculation for image adjustment
- 4. Optimisation Limit possible radius, thus limit search space Couple of voting locations served by one block Automatic adjustment of thread and block numbers to keep load higher - more efficient When pupil is found, iris search space only around the same centre Avoiding host-device and device-host memory copying Using shared memory
- 6. Results - Timing Time measurements of the whole process including image pre-processing Image size C++ CPU CUDA Speed up 1024x768 32.490 sec 3.947 sec 8.23x 800x600 17.161 sec 2.054 sec 8.35x 640x480 8.701 sec 1.120 sec 7.76x 500x375 4.202 sec 0.589 sec 7.13x 350x263 1.461 sec 0.293 sec 4.99x 250x188 0.625 sec 0.175 sec 3.57x MATLAB with 640x480: over 32 sec
- 7. Results - Timing Iris detection CPU vs CUDA 40.000 30.000 Time, sec 20.000 10.000 0 250x188 350x263 500x375 640x480 800x600 1024x768 Image size CPU CUDA
- 8. Hardware Used HP All-in-One 200-5300ch Processor: Intel Core i3-550 (3.2GHz), 4MB RAM: 2x 2048MB DDR3, 1333MHz GPU: nVidia GeForce G210 with 512 MB (low-end)
- 9. Comments and Future Improvements Algorithm should work faster on a better GPU Main bottleneck - AtomicAdd function Get rid of any CPU based pre-processing Make the whole process purely on GPU, no memory copying involved except for the result Automatic parameter adjustment to ensure good results on variety of pictures