红外与激光工程, 2018, 47 (4): 0404001, 网络出版: 2018-09-19   

基于一类支持向量机的盲元检测方法

One class support vector machine used for blind pixel detection
张东阁 1,2,*傅雨田 1,2
作者单位
1 中国科学院红外探测与成像技术重点实验室, 上海 200083
2 中国科学院上海技术物理研究所, 上海 200083
摘要
利用无监督学习的一类支持向量机(One Class Support Vector Machine, OCSVM)和随机场景图像序列, 构造滚动更新的像元分类模型, 实现红外焦平面盲元的在线检测。根据正常像元和异常像元数量和灰度特征的差异, 以随机图像序列作为输入数据, 使用OCSVM建立单一类别的像元分类模型, 灰度变化的像元归为一类, 其他像元不属于此类。由于随机图像序列的滚动更新, OCSVM模型及支持向量也随之更新。统计支持向量的频次, 高频次支持向量对应的像元聚为一类, 即为异常像元。以320×256中波红外图像序列为例, 说明了OCSVM模型进行盲元检测的过程, 检测结果与黑体定标的结果一致。基于随机场景和OCSVM模型的盲元检测方法摆脱了定标黑体的约束, 提高了盲元检测的灵活性。
Abstract
One class support vector machine(OCSVM) was applied to classify the pixels of the infrared detectors, and it can detect the blind pixels by the random scenes. The blind pixel detection algorithms were reviewed in the beginning, and the imbalance distribution of the normal pixels and blind pixel was discussed in the following. The infrared image sequence was used to set up the OCSVM models and calculate the super sphere parameters, when the support vectors were represented by the Lagrangian coefficients. The OCSVM was an unsupervised method to cluster the pixels by the changing gray level and the random scenes. The super sphere model built by OCSVM would be refreshed by the updating image sequence, while the Lagrangian coefficients of the support vectors were recorded, so the blind pixels could be eventually classified by the statistic results of the preceding coefficients series. The mid-wave infrared 320×256 image sequence was taken as an example to illustrate the proposed method, and it got the same results as the black body calibration. It could conclude that the OCSVM used for the online modeling of the blind pixel detection of the infrared detectors is adaptive and self-refreshing, and it could improve the efficiency of the infrared system test.中国科学院上海技术物理研究所创新基金(2014-CX25)

张东阁, 傅雨田. 基于一类支持向量机的盲元检测方法[J]. 红外与激光工程, 2018, 47(4): 0404001. Zhang Dongge, Fu Yutian. One class support vector machine used for blind pixel detection[J]. Infrared and Laser Engineering, 2018, 47(4): 0404001.

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!