针对复杂情况下海上舰船目标单波段特征识别能力不足的问题, 研究可见光、 中波红外和长波红外三波段特征图像融合技术, 重点解决图像融合方法中存在的算法耗时和融合策略选择的问题, 提出了一种新的基于区域协方差矩阵的多波段特征级融合方法, 针对可见光图像和红外图像分别设计11维和5维特征向量, 协方差矩阵可以将多个特征进行融合, 既保证了不同目标之间的区别性, 同时又减小计算量。 该方法首先利用显著性检测, 快速定位图像中的目标区域, 然后, 针对不同波段图像设计的特征向量定义协方差阵的距离计算公式并进行匹配, 通过对图像的一次遍历操作获得积分图像, 在协方差计算时达到快速计算的目的, 最后利用k-阶最近邻算法对多种舰船目标进行分类识别。 利用实拍的3 400余张三波段舰船目标图像作为测试数据。 实验主要分为两部分, 首先对比单波段和三波段融合识别的识别率, 验证所提出的融合方法具有更广的应用范围; 然后, 在计算效率上对比多种传统的像素级方法, 验证采用的特征级融合在计算时间上的优势。 实验结果表明, 该方法可达到95.1%的识别率, 单帧计算耗时约为0.5 s, 在实时性和检测率方面都有明显提高。

Aiming at solving the problem of inadequate ability of recognizing single band warship targets on the sea in complicated situations, image fusion technology for three band feature levels of visible light, short wave infrared and medium wave infrared has been studied. Problems of time-consuming and fusion strategy existed in image fusion have been solved as top priorities. A new multi-band fusion method based on regional covariance matrix has been proposed in this paper. The advantage is that it can fuse several related features naturally through low dimensional vector. This paper has designed 11-dimensional and 5-dimensional feature vectors for visible light image and infrared image respectively, which not only ensures the differences among different targets but also decreases the calculation. Saliency detection has been adopted firstly in this paper to position the target area in the image quickly; and then, distance calculation formula of covariance matrix has been defined for the feature vectors of different band image designs and matching has been made in the following; integral image has been attained through one traversal operation for the image; purpose of quick calculation has been realized in the calculation for covariance; finally, K-nearest neighbor (KNN) algorithm has been adopted to make classification and recognition for various warship targets. Over 3 400 images of three-band warship target have been used as testing data. The experiment mainly includes two parts: first, comparing the recognition rate of single-band and three-band fusion recognition and to verify that the fusion method proposed in this paper has more extensive application range. Second, comparing several traditional pixel level methods in the calculation efficiency to verify the advantages of feature fusion adopted in this paper in calculation time. Experimental results have shown that the recognition rate of this method can reach 95.1% and time for single frame calculation is about 0.5 s, which has made distinctive improvement in both real time and detection rate.