一种基于弱监督学习的声图小目标快速检测方法*

小目标检测是声呐图像理解中最引人瞩目,同时又极具挑战性的任务之一。该文基于离散余弦变换和K-近邻聚类,提出了一种快速检测方法。离散余弦变换用于生成图像的指纹,是原始图像在二维频域的一种稀疏表达;改进的K-近邻模型对于带有标签数据的需求量相对较低,提升了算法的处理效率和对弱监督场景的适应性。经试验验证,该方法可在准确率和召回率之间达到一个恰当的平衡点,同时在实时成像的合成孔径声呐图像小目标检测中,获得了较为可靠的结果。

A fast weak supervised detection method of small objects in sonar imagery

Detection of small objects is one of the most attractive and challenging tasks in the comprehension of sonar imagery. In the paper, a fast detection method is presented under a framework of Discrete Cosine Transform (DCT) and K-Nearest Neighbor (KNN) Clustering. DCT is used in the generation of image fingerprint, which contributes a certain spectral sparseness to the original image; and the modified KNN model provides efficiency with a relatively low demand of labeled data. It is shown in a series of experiments that the method we propose can reach a compromise of precision and recall rate, and achieve considerably reliable detection result on synthetic aperture sonar (SAS) images in real time imaging.