侧扫声呐图像特征提取改进方法

Improved Feature Extraction Method for Side Scan Sonar Images

由于侧扫声呐图像对于研究海底状况有着重要的作用,因此常采用点特征和线特征提取来分析声呐图像的特点。其中点特征提取分析采用尺度不变特征变换(SIFT)算法,直线特征提取采用Hough变换和线段检测(LSD)算法。文中针对Hough变换在提取直线特征时因基本空间转换方式存在不足而使得提取的线特征不理想的问题,提出了一种效果更好的空间转换方式,该方法需要建立2个图像边缘掩码矩阵,一个保持原状用于寻找直线特征,另一个根据边缘点转化情况相应减少对应边缘点,这种方法的设计可以保证属于直线的特征点都能参与到直线特征的形成中;针对LSD算法中检测到的直线特征会因相交而出现断裂的问题,设计了一种断裂线特征拟合方法,利用角度、位置等条件将被判定为断裂的直线特征进行拟合,减少因直线特征断裂造成直线过短过多的问题,然后通过长度阈值筛选出合适的直线特征。通过优化前后的方法对侧扫声呐图像进行直线特征提取验证,证明了所改进的Hough变换可以更好地提取到线特征,而优化后的LSD算法也减小了线特征断裂的影响。同时对于待处理的声呐图像,点特征提取数量较多,不能很好地体现图像内容特点,而改进之后的线特征可以提取出图像内容的轮廓,信息丰富,能更好地表现了图像的内容。以上改进方法可为有关线特征提取研究提供参考。

To solve the problem in feature extraction of side scan sonar images that the line feature extracted by Hough transform is not satisfactory due to the deficiency of the basic space conversion method, a better space conversion method is proposed. This method needs to establish two image edge mask matrices - one for maintaining the original shape to find the straight line features, and the other for reducing the corresponding edge points according to the edge point transformation. This method can ensure that the feature points belonging to the straight line participate in the formation of linear features. However, the linear features detected in the line segment detection(LSD) algorithm will be broken due to intersection, hence a method of fitting the broken line features is designed, in which the straight line features being determined as ‘broken’ are simulated via angle and position so as to deal with the trouble of having too short or too much straight lines due to the fracture of the linear features, and then the appropriate straight line features are selected according to the length threshold. The linear feature extraction of side scan sonar images verifies that the improved Hough transform can extract the line features better, and the optimized LSD algorithm can reduce the influence of line feature fracture. For the sonar image to be processed, the number of point feature extraction is large, which can not reflect the characteristics of the image content well;and the improved line feature can extract the outline of the image content, and the information is rich, which better expresses the content of the image. It is provide reference for relevant line feature extraction research.