摘要
在基于SAR图像的海上溢油检测中,识别的效率与准确率是关键。纹理特征是人类专家能够较好的判别SAR图像中油膜,类油膜以及海水的一个重要依据。本文算法一方面融合灰度共生矩阵与Tamura特征,直接对SAR原始图像进行特征提取,避免了对图像进行分割、降噪等预处理,提高了识别算法的可行性与识别效率。另一方面,应用深度信念网络(DBN)的分类方法,可以很好地解决溢油检测中小样本分类的问题,并且模仿人类感知系统高效准确的表示信息、获取本质特征。本文应用人类感知的思想对油膜、类油膜以及海水这3类样本进行分类识别。通过实验确定了DBN中利于分类的关键参数值。本算法对原始SAR图像中3类样本的识别准确率达到90.36%,具有较好的实用价值。
In the study of marine oil spill detection based on SAR images, identification efficiency and accuracy are the key issues. A human expert can determine the oil film,look-alikes, and sea surfaces more easily from texture features, so texture features are an important resource. On the one hand, our algorithm merges Tamura and GLCM features with the original SAR image, and extracts features from SAR image directly. This is an attractive solution for oil spill detection using small samples. This approach also avoids image segmentation, denoising preprocessing, and improves the feasibility and efficiency of identification recognition algorithms. On the other hand, we applied the deep belief network （DBN） to mimic the human perception system＇s efficient and accurate representation of in- formation to get the essential features. We use the idea of human perception to classify the three kinds of samples （oil film, look-alikes and sea surface）. Through experiments; we determined key parameter values in the DBN. Recognition accuracy of this algorithm on three kinds of samples of the original SAR image reached 90.36%, and it has good practical value.
出处
《武汉大学学报（信息科学版）》
EI
CSCD
北大核心
2016年第3期395-401,407,共8页
Geomatics and Information Science of Wuhan University
基金
国家海洋局海洋公益性行业科研专项基金(201305026-3)
上海市科学技术委员基础研究重大重点项目(08240510800)~~
