Extreme Contour Point를 이용한 회전 이미지 각도 검출 알고리즘의 FPGA 설계 : FPGA Implementation of Novel Rotation Angle Detection Algorithm using Extream Contour Point

다국어 초록 (Multilingual Abstract)

In this paper, we propose a novel algorithm with fast speed and low cost using Extreme Contour Point (ECP) to detect the angle of rotated high definition (HD) images is proposed. it is designed by rotation invariant feature of overall images that can be applied to correct the rotated images for real time application. There are two advantages for ECP algorithm, namely compatibility to the images in pre-processing by using Sobel edge process for pattern recognition. While the other one is its simplicity for the rotated angle detection with cyclic shift of two 1⨉n matric sets between original image and rotated image.
The initial input images can be concisely assembled in its edge as their significant features which are being represented by Sobel edge based on gray scale form. And, after a binary conversion of the extracted image, an additional filtering needs to make a rotation invariant feature is called as the ECP. There are three schemes on the ECP according to extraction method and its inherent variables such as ECP(Ɵ), ECP(r2,Ɵ) and ECPAM(r2,Ɵ). ECP(Ɵ), menas that the feature of its components depends on only angle Ɵ, However, could be vulnerable to a noise which represents in changed rotation axis from the original rotation axis. In real time process of moving camera, the rotation axis might be possible to infinitesimally shift for frame to frame. On this wise, ECP(r2,Ɵ) is designed to form relatively non-sensitive about infinitesimal variation of rotation axis by comparing ECP(Ɵ), that is expressed not just only the features of angle Ɵ but the square of radius r2. Lastly, ECPAM(r2,Ɵ) is improvement version in terms of accuracy about rotated angle, is derived from ECP(r2,Ɵ) with each components matched to angles from 0 degree to 360 degree.
Through the three ECP scheme, rotation angle detection is evaluated with various 100 samples, and its detectable angle is limited to 10 degree, 5 degree and 20 degree for ECP(Ɵ), ECP(r2,Ɵ) and ECPAM(r2,Ɵ), respectively. Especially, ECPAM(r2,Ɵ) uses look up table (LUT) to match the angle. Moreover, it is additively designed to optimize LUT by reducing the extraction radius from r=57 to r=45. This result gives LUT reduction effect of 37.7% in comparison with r=57. And then, relative speed is simulated by using Matlab R2012a based on Intel Core i5 3.3GHz CPU environment, that the process speeds of ECP(Ɵ), ECP(r2,Ɵ) and ECPAM(r2,Ɵ) are 7ms, 15ms, and 20ms, respectively. While rotation matrix method based on CORDIC is 143ms.
Finally, through optimized hardware design based on field programmable gate array (FGPA) about proposed three ECP versions, it was evaluated by XC7Z020 xc7z020-3clg400 FPGA board by using xilinx 14.2 tool. Coordinate rotation digital integrated computation (CORDIC), is a typical algorithm to estimate rotated angle between points, is used as benchmark algorithm. By comparing our proposed algorithms and the benchmark, the benchmark, ECP(Ɵ), ECP(r2,Ɵ), and ECPAM(r2,Ɵ) are confirmed to have similar Max. operation frequency of 242MHz, 280MHz, 229MHz, and 255MHz, respectively. However, it is verified to have high performance result in terms of the hardware cost, ECP(Ɵ) is much better than the benchmark with low cost reduction of resister and LUT of 0.08% and 0.5%, respectively. And ECP(r2,Ɵ) and ECPAM(r2,Ɵ) are also much better than the benchmark with low cost reduction of resistor and LUT of 0.10% and 1.00%, and 0.18% and 1.48%, respectively.
From this experiment result, the proposed ECP algorithms are proven to be feasible enough to in real time application. And, ECP(Ɵ) is highly efficient with high speed and much low cost in the system is not changed to rotation axis. While ECPAM(r2,Ɵ) is optimized to the system with very improved accuracy.

목차 (Table of Contents)

• I. Introduction 1
• II. Background of the research 6
• 1. Benchmark algorithm (CORDIC) 6
• 2. Sobel edge process 8
• 3. Extraction of Sub-region for rotation invariant feature 10
• III. Theoretical configuration of extreme contour point 14
• 1. Theta Ɵ dependence ECP(Ɵ) 14
• 2. Square of Radius r2 and theta Ɵ dependence ECP(r2,Ɵ) 18
• 3. Angle matched ECPAM(r2,Ɵ) 21
• IV. Software simulation result 24
• 1. Rotation angle detection result 24
• 1-1. Theta Ɵ dependence ECP(Ɵ) 26
• 1-2. Square of Radius r2 and Theta Ɵ dependence ECP(r2,Ɵ) 28
• 1-3. Angle matched ECPAM(r2,Ɵ) 33
• 2. Relative process speed measurement 44
• V. Hardware Implementation 46
• 1. Hardware architecture 46
• 1-1. Input line buffer module 47
• 1-2. ECP process module 47
• 1-2-1. ECP(Ɵ) filtering module 47
• 1-2-2. ECP(r2,Ɵ) filtering module 48
• 1-2-3. ECPAM(r2,Ɵ) filtering module 50
• 1-3. Sum of Absolute difference module 52
• 2. FPGA experimental result 53
• 3. Synthesis result 64
• VI. Conclusion 66
• References 68
• 국문 요약 72