Efficient Architectures for Segmentation of Endoscopic Images in Micro-Robotic Auto Navigation Systems

This paper presents novel techniques and their architectural translations for real-time image segmentation of endoscopic images, which are required for micro-robotic auto navigation systems. The proposed technique is based on a two-step process to segment the lumen regions from endoscopic images. In the first step, an adaptive progressive thresholding technique based on Otsu's method is employed to obtain a preliminary region of interest of the lumen region. A novel architecture for the between-class variance computation of Otsu's method is presented to meet the real-time requirements of the system. The proposed implementation employs binary logarithmic computations to eliminate the complex divisions and multiplications in Otsu's procedure. In the second step, an Iris filter is employed to enhance the boundary of the region of interest to facilitate an accurate detection of the lumen region. An architecture based on the coordinate rotation digital computer is proposed to simplify the complex computations of trigonometric functions required by the Iris filter operation. Software simulations demonstrate that the proposed technique requires a significantly smaller number of iterations to obtain an accurate segmentation result as compared to a previously reported method. In addition, synthesis results on the field-programmable gate array show that the proposed architectures can achieve high performance with low hardware resource utilization.