一种基于红外图像特征融合的高温铝液模拟泄漏监测算法

Simulated leakage monitoring algorithm for high-temperature molten aluminum based on the infrared image feature fusion

高温熔融铝液泄漏是炼铝工业最严重的危害之一。借助FLIR A310红外热像仪开展试验研究,搭建了高温铝液泄漏模拟试验平台,建立模拟泄漏标准数据集,提出了一种基于红外图像特征融合的高温铝液模拟泄漏监测算法。利用HOG和LSS描述子分别提取图像梯度边缘和相似形状等几何特征并进行特征向量融合,弥补HOG单一特征检测的不足,将融合的特征向量送入训练好的RBF核函数支持向量机分类识别。融合算法试验性能最优,测试集的查准率、查全率和F1分别为94.31%、93.52%和93.91%。LSS的高维特征向量影响监测时间,步长为15、维度为5280时,可以实现特征降维与识别准确率的平衡。最后,探究了特征融合描述子对相机距离的敏感性,3.5 m可实现最佳识别效果;200张图片的监测时间稳定在48.59 s左右。

This paper in intended to make a simulated research to establish the standard data set of the simulated leakage and build up a molten aluminum leakage situation with the hot water by getting a total set of 1800 positive and negative samples with the typical features.And,then,we have built up a simulation experiment platform for the high-temperature molten aluminum leakage via the FLIR A310 thermal imaging camera,so as to solve the problem of high-temperature molten aluminum leakage.And,furthermore,we have also proposed a simulation leakage monitoring algorithm for high-temperature molten aluminum based on the infrared image feature fusion.What is more,we have succeeded in extracting the image gradient edge,the similar shape feature,respectively,and the feature vector fusion,to make up for the shortage of the single feature detection of HOG by using the histogram of the oriented gradient(HOG)and the local self-similarity descriptor(LSS).And,next,the fused feature vectors have been transmitted to a trained RBF kernel support vector machine(SVM)for classification.The algorithm we have proposed here can be said of great help to achieve the excellent experimental results,with their precision,recalling and F1 of the test data set being equal to 94.31%,93.52%and93.91%,respectively.Hence,the high-dimensional feature vector of the local self-similarity descriptor can also influence the monitoring time.Thus,with the gradual increase of the step size,the feature granularity tends to be getting coarser and coarser,but both the precision and the recall of the test data set are gradually getting to decrease.And,when the step size is set to15 and the dimension is 5280,the balance between the feature dimension reduction and the recognition accuracy can be gradually achieved.And,in so doing,we have finally made an exploration of the sensitivity of the feature fusion descriptor to the camera distance.The experimental results show that the best recognition performance can be realized at 3.5 m,corresponding to the actual camera distance of about 10.5 m,and the precision,recall and F1 can be expected to reach 96.73%,96.50%and96.61%,respectively.In addition,the total monitoring time of200 pictures can be shortened to as much as around 48.59 s,with no change of the camera distance.