基于贝叶斯统计推理的复杂场景边缘检测

复杂场景图像存在大量的噪声和纹理干扰,传统边缘算子的检测效果不理想.为此,文中提出一种基于贝叶斯统计推理理论的多信息融合边缘检测算法.该算法融合了梯度算子、拉普拉斯算子以及两级均值比率(ROA)算子的输出响应;通过最大类间属性互信息对特征属性进行最优离散化;利用非参数直方图方法估计类概率密度函数;并通过贝叶斯风险最小化原则实现边缘检测.试验中该方法的Bhattacharyya误差界为0.093,工作特性曲线下的面积(AUC)达到0.958,证明了该算法的有效性.与经典算子检测结果的比较也表明,该算法能够有效地克服图像中的噪声和纹理干扰.

Edge Detection of Complex Scenes Based on Bayesian Statistical Inference

The traditional edge detectors are inefficient in the image detection in complex scenes due to the disturbances of noise and texture.In order to solve this problem,a multi-information fusion edge detection algorithm based on Bayesian statistical inference theory is proposed.This algorithm fuses the output responses of four operators,such as the gradient operator,the Laplacian operator and the ratio of average(ROA) operators at two scales,achieves the optimal discretization for the continuous attributes of feature vector by maximizing the class-attribute mutual information,employs the nonparametric histogram method to estimate the class-conditional probability density functions,and adopts the principle of Bayes Risk Minimization to complete the edge detection of new images.Experimental results show that the proposed algorithm is feasible,with a Bhattacharyya error bound of 0.093 and area under the receiver operating characteristic carve(AUC) of 0.958.The comparison of the detection results obtained respectively from the proposed algorithm and from the classical detectors also shows that the proposed algorithm is robust to the noise and texture in images.