SAR图像ROEWA边缘检测器的改进

Improved ROEWA edge detector for SAR Images

针对指数加权均值比(ROEWA)边缘检测算子无法准确确定边缘位置和边缘方向的问题,对其表达式进行了改进。重新定义了反转的、带符号的、归一化最小指数加权均值比(IROEWA),作为边缘强度指标,来量化描述边缘的跃变程度,并在此基础上准确计算出边缘的方位。同时将改进的非极大值抑制算法应用到边缘检测流程中,对IROEWA得到的边缘强度图进行后处理。针对自然SAR图像和仿真SAR图像的实验都表明,改进算法能够得到比ROEWA加分水岭分割的原始算法更好的边缘检测结果。利用接收者操作特征(ROC)曲线,对改进算法进行量化评价,其曲线下面积(AUC)可达0.97570,非常接近理想的检测器;在其ROC曲线的最优点处,检测率可达0.95232,而虚警率仅为0.00214左右。

This paper reports on a method to improve the Ratio Of an Exponentially Weighted Averages（ROEWA） edge detector, so that the improved edge detector can accurately determine the positions and directions of edges for Synthetic Aperture Radar（SAR） images. We attempt to build an optimal edge detector for SAR images to obtain better results of edge detection. The edge strength index is redefined as an inverted, signed, and normalized minimum ROEWA（IROEWA）, which is utilized to quantitatively describe the phase step of edges. A new method that accurately calculates edge direction is developed based on the edge strength map from IROEWA. We can obtain the possible values of edge directions in this manner, which continuously distributes from 0 degrees to 180 degrees. Therefore, we must improve the Non-Maximum Suppression（NMS） algorithm, so that it can process sub-pixels. Finally, the improved NMS algorithm is also added into the edge detection workflow. This improved edge detection algorithm is called IROEWA NMS. We conducted two experiments for IROEWA NMS： one employed nature SAR images, whereas the other adopted a simulation SAR image. Experiment results show that the IROEWA NMS outperforms the original ROEWA with watershed thresholding. The IROEWA operator is faster than the ROEWA operator under the same conditions. We applied a Receiver Operating Characteristic（ROC） curve to evaluate the IROEWA NMS and determined that its Area Under the Curve（AUC） is 0.97570; thus, it approximates the ideal optimal detector. The detection rate at the position of the optimal point in the ROC curve of the IROEWA NMS is as high as 0.95232, whereas the false alarm rate is as low as 0.00214. The IROEWA NMS exhibits suitable performance on both the detection and false alarm rates. It has significant application value in several fields, such as the segmentation and edge detection for SAR images.