基于水平集和支持向量机的图像声呐目标识别

为了能实现水下遇险目标的精确定位,首先要对声呐所获取的图像进行目标识别。利用水平集法获得水下声呐图像中目标轮廓后,提取目标轮廓的7个不变矩作为特征矢量,并将获取的不变矩特征输入到已经训练好的支持向量机中进行识别,从而得到识别结果。所使用的识别方法综合了基于水平集提取轮廓的长处,不变矩的位移、尺度、旋转不变性的特点和支持向量机在小样本、非线性模式识别中的独特优势。实验结果表明:该方法对高分辨率图像声呐具有较高的识别率和较低的误判率,对原始声呐图像的目标识别率高达99%,对加入方差为0.09的高斯噪声的声呐图像的目标识别率可以达到97%。

Level-set and SVM based target recognition of image sonar

In order to locate the submarine target in danger precisely,the first step is to recognize the target from the sonar image.The proposed method in this paper extracts target contours using level-set,calculates invariant moment of the contours as the feature vectors and then inputs the feature vectors to a support vector machine(SVM) that has been trained to obtain the recognition results.This algorithm combines the merits of the level-set,i.e.the invariability of shifting,rotation and scaling and the unique advantage of the support vector machine in small samples and nonlinear pattern recognition.Experiment results show that the proposed method has higher recognition rate and lower false-positive rate: the recognition rate of the original image achieves 99%,and the recognition rate of the sonar image with Gaussian noise whose variance is 0.09 reaches 97%.