반복 연산 기반 Otsu 방법을 활용한 에지 검출 가속기

초록

Edge detection is a key technology that significantly enhances the performance of vision-based systems by extracting object boundaries,and its importance is increasingly recognized in various real-time applications. To generate an edge image, selecting a threshold todistinguish between edge and non-edge regions is essential. The Otsu method is an effective technique for calculating the optimal thresholdthat divides an image into two classes. In particular, the Otsu method automatically selects a suitable threshold for each image, enablingclass classification in environments where images change in real-time. However, implementing the Otsu method in hardware typicallyinvolves the use of logarithmic approximation techniques, w hich improve computational efficiency but increase hardw are resourceconsumption. This paper proposes an edge detection accelerator that enhances hardware resource efficiency by introducing ahistogram-based iterative computation approach to the process of finding the maximum between-class variance in the Otsu method. Performance evaluation using 100 images from the ImageNet dataset revealed that the proposed method achieved an absolute error of5.04 compared to the theoretical value in Otsu threshold calculation, and the binary classification using the computed threshold achievedan accuracy of 97.88%. Also, synthesis results show that the proposed hardware architecture utilizes 9,442 slice LUTs and 5,060 sliceregisters in the Otsu module. The proposed edge detection accelerator reduces hardware resource consumption while maintaining highcomputational accuracy compared to conventional logarithmic approximation techniques, making it suitable for various real-time imageprocessing applications.

키워드