| Summary: | 碩士 === 國立中興大學 === 電機工程學系所 === 102 === The research of this study focuses on two digital image process issues. One of them is dynamic scenes object tracking, which adopts a feature points matching based tracking scheme, another one is blurred image recovery, which uses digital image process techniques to recover out of focus or motion blurred images. In object tracking, we apply a features points matching scheme to track the object. This method shows balanced performance between algorithm complexity and tracking efficiency. Even operating in complex backgrounds, this method still has good tracking performance. In order to support real-time tracking in actual environments, the kernel of the object tracking system is realized on a VeriEnterprise VEX-S150 FPGA platform. The FPGA platform is further integrated with a host PC equipped with a camera to form the entire system. This shows a typical example of HW/SW co-design. The host PC is in charge of the video capture, display and communication with the FPGA platform. In particular, the captured video data is passed to the FPGA for hardware acceleration under the control of host PC. After the tracking process, the result is passed from FPGA to host PC for display. The proposed system is capable of performing the tracking at frame rate of 31 fps, which is 10 times faster than a pure software implementation. In blurred image recovery, this study adopts a new point spread function estimation scheme to recover blurred images. This scheme exploits the edge-spread phenomenon, which is a unique feature of blurred images, and estimate edge spreads from different edges automatically. Since human eyes’ perception on whether an image is sharp or not, is based on the sharpness of the edges, so we just focus on processing the edges of the image in the deblurring process. This can also reduce the computing complexity significantly. The proposed scheme begins with transforming the image to the frequency domain to check the existence of motion blur. It then records the edge spread functions from every orthogonal direction of the edges. The obtained edge spread functions are next converted to line spread functions, where approximate Gaussian models are chosen to reconstruct the PSF model. If the blurred image had motion blur, the motion blur model would be added to PSF model. After PSF model estimation, the deconvolution process can be performed by using Wiener Filter. Since every edge can have its unique PSF model, the proposed scheme is particularly useful in handling blurred images with multiple PSFs. Finally, Smoothening process of ringing artifacts, which often appear around the edges of the image, is applied. A median filter is also employed to reduce the noise effect of the Wiener Filter. Based on the experimental results of motion blur, the proposed scheme can achieve 2 times better than the conventional deblurring schemes such as exhaustive method. Even based on the experimental results of out-of-focus, the proposed scheme can achieve 90 percentage of efficacy of exhaustive method, but requires less than half of the time.
|
|---|
