光谱学与光谱分析, 2019, 39 (3): 698, 网络出版: 2019-03-19   

多波段红外图像的海面舰船目标检测

Research on Detection of Ship Target at Sea Based on Multi-Spectral Infrared Images
作者单位
1 海军航空大学, 山东 烟台 264000
2 烟台大学光电信息科学技术学院, 山东 烟台 264000
摘要
现有的基于单个红外宽波段的海面舰船目标探测系统在面对复杂海天背景、 岛岸背景、 恶劣天气、 亮带干扰或诱饵弹干扰等情况时, 系统的探测率、 虚警率、 探测距离等性能指标均会受到严重的影响; 为此, 开展了基于多波段红外图像的海面舰船目标检测方法的研究。 通过中波红外多波段数据采集系统实际采集107组五个中波红外波段的图像; 波段1—5分别为3.7~4.8, 3.7~4.1, 4.4~4.8, 3.7~3.9和4.65~4.75 μm; 对多波段图像进行手动标注构建样本数据集, 其中, 正样本舰船目标298个, 负样本非舰船目标353个。 对于多波段红外图像, 首先进行PCA降维并采用选择性搜索算法生成初始目标候选区域; 针对候选区域中存在大量明显的非舰船目标区域的问题, 利用积分图像计算候选区域的局部对比度, 依据红外舰船目标的几何和灰度特征从初始目标候选区域中筛选出舰船目标可能性大的区域作为舰船目标候选区域。 然后对舰船目标候选区域进行拓展以融入局部上下文信息, 对于候选区域对应的5波段红外图像, 分别提取每个波段图像的稠密SIFT特征, 并将128维SIFT特征向量降为64维, 融入SIFT特征的空间和波段位置分布信息得到新的特征向量, 基于高斯混合模型对候选区域的特征向量集合进行编码融合得到舰船目标候选区域的费舍尔向量表示, 最后利用线性SVM分类器识别出舰船目标。 对多波段图像进行舰船目标候选区域生成实验, 所提出的基于红外舰船目标的几何和灰度特征的约束方法可以有效地克服选择性搜索算法的不足, 从初始目标候选区域中快速定位出舰船目标候选区域, 对25组多波段图像进行实验, 舰船目标候选区域生成的整体耗时为0.353 s, 定位舰船目标区域耗时0.005 s。 对100个正负样本进行目标识别测试, 所提出的目标识别算法融合了目标的多波段图像特征信息, 通过引入费舍尔向量挖掘了多波段图像梯度统计特征的深层次信息, 算法的识别率达到了0.97, 显著高于单波段红外图像的目标识别率。 对25组多波段图像进行舰船目标检测实验, 所提出的舰船目标检测方法能够在海天背景、 岛岸背景以及亮带干扰等不同场景下完成海面舰船目标的检测工作, 舰船目标定位准确, 舰船目标召回率达到了0.95, 每组多波段图像的平均检测耗时为1.33 s。 研究结果表明, 充分考虑海面舰船目标在红外图像中与局部海洋背景的辐射差异以及有效地融合舰船目标在多个红外波段图像中的辐射特征, 可以增强舰船目标的可分性, 提高舰船目标的识别率以及检测率, 为基于多波段红外图像的海面舰船目标检测提供了新的技术支持。
Abstract
When facing complex sea-sky background, island-shore background, bad weather, bright waves or decoys interference and other complex conditions, the detection rate, false alarm rate, detection distance or other performance indicators of the existing ship target detection system based on a single wide-wave infrared image will be affected. Considering the above problems, the detection method for ship target at sea based on multi-spectral infrared images was studied in this paper. Through the data acquisition system for multi-spectral infrared images, 107 groups of 5 medium-wave infrared images were collected actually. The spectrals from 1 to 5 were 3.7~4.8, 3.7~4.1, 4.4~4.8, 3.7~3.9 and 4.65~4.75 μm respectively. The sample data set was constructed by annotating the multi-spectral images manually, which was made up of 298 ship targets and 353 non-ship targets. Firstly, PCA transform was adopted to reduce the dimension of multi-spectral infrared images and selective search algorithm was adopted to generate the initial target candidate regions. In order to solve the problem that there are too many obvious non-ship target regions, the integral image was used to calculate the local contrast of the initial candidate regions and the ship target candidate regions were located according to the geometrical and grayscale features of infrared ship target. Secondly, each ship target candidate region was extended to incorporate the local context information. For the 5 spectral images corresponding to each ship target candidate region, dense SIFT feature of each spectral image was extracted. PCA was applied to SIFT feature, reducing its dimensionality from 128 to 64. Then the spatial and spectral position distribution information of each SIFT feature was added to the feature vector. Based on the Gaussian mixture model, the feature vectors of each candidate region were encoded to Fisher vector representation. Finally, linear SVM classifier was used to recognize ship targets. Experiment of the generation of ship target candidate regions showed that the proposed constraint method based on geometrical and grayscale features of infrared ship target can effectively overcome the shortcomings of selective search algorithm and quickly locate the ship target candidate regions from the initial target candidate regions. Experimental results on 25 groups of multi-spectral images showed that the generation of ship target candidate regions takes 0.353 s totally, while locating the ship target candidate regions takes only 0.005 s. Test of target recognition on 100 positive and negative samples showed that the recognition rate of the proposed target recognition algorithm reached 0.97, which is significantly higher than the target recognition rate based on single-wave infrared image. The proposed target recognition algorithm integrates the feature information of the multi-spectral target images and applies Fisher vector to extract the deep layer information in the gradient statistical features of the multi-spectral target images. Experimental results on 25 groups of multi-spectral images showed that the proposed ship target detection method can detect the ship targets at sea in different scenes such as sea-sky background, island background and bright waves interference. The locations of the ship targets are accurate and the ship recall rate reaches 0.95. The average detection time of each group of multi-spectral images is 1.33 s. The study results showed that with considering the radiation difference between the ship target and its local ocean background in the infrared image and the effective fusion of the radiation characteristics of ship target in multi-spectral infrared images, the divisibility of ship target can be enhanced, which results in the improvement of recognition rate and detection rate of ship target. This study provides new technical support for ship target detection based on multi-spectral infrared images.

仇荣超, 娄树理, 李廷军, 宫剑. 多波段红外图像的海面舰船目标检测[J]. 光谱学与光谱分析, 2019, 39(3): 698. QIU Rong-chao, LOU Shu-li, LI Ting-jun, GONG Jian. Research on Detection of Ship Target at Sea Based on Multi-Spectral Infrared Images[J]. Spectroscopy and Spectral Analysis, 2019, 39(3): 698.

本文已被 3 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!