Detecting light loss NDT industrial -X small sample
- Small sample: insufficient amount of data, there is not enough data to support;
- X-ray detection of loss: as an industrial non-destructive testing of industrial technology. Continuously irradiated by X-ray object to be detected for some time, and the intensity of X-rays pass through various parts of the object to be detected by the recording film. Where high penetration strength, the film portion corresponding to the gradation value is relatively large, relatively bright; penetration strength weak areas, the film portion corresponding to gradation value is relatively small, relatively dark. By observing the gray value of the X-ray film mulberry object to be detected can be inferred whether the internal structure of the expected production.
research method
To solve the 小样本数据下工业无损检测图像中的损伤定位问题proposed 一种结合主动学习的多尺度多特征融合的损失检测方法, combined 目标识别的方法to reduce the classification model based on a sliding window to bring the false positive rate, to further enhance the effect of the loss of positioning model.
- 1. Active learning: active learning through policy, without affecting the positioning accuracy of the model, reduce labeling costs model training needs.
- 2. Multi-features: enhanced by the feature enhancement method damage features in the image, and separated by a clustering method in X-ray image of the weld area and weld steel region, and fuses the various features described above, the information obtained richer multi-feature fusion data.
- 3. Multi-scale discrimination: the use of a smaller box to locate the loss of positioning, enhance the accuracy of positioning loss model.
Multi - feature in the active learning method of positioning the frame loss
By improving CLAHE method proposed G-CLAHE method for enhancing X-ray images loss gradient, contrast characteristics and solve the problem of uneven distribution of gray values. X-ray image information of the position of the weld followed by extraction clustering way, by this empirical method was added into the manual inspection Injury Model. After the original image, G-CLAHE feature enhancing clustering results combine to form a three-channel data richer multi-feature fusion solve due to the different dimensions of the data can not be ImageNet use of the model pre-trained, while Since the data richer features to further enhance the classification performance of the model. On the other hand, this paper presents a multi-scale discriminant model, so the model to get more information while using less space positioning box to locate the damage, damage detection improves the performance of the pixel level. Finally, active learning methods, to ensure the accuracy of positioning damage, while reducing the amount of data labeling costs.
1. The multi-feature fusion
X-ray images is a single channel TIFF format data, and the current image data set of common data bit three-channel RGB format. 已有的方法通常是将每张单通道的图像数据复制为三份，合并成一个三通道的图像数据. Here 利用非线性灰度增强改进了CLAHE(限制对比度的自适应直方图均衡方法)，得到G-CLAHE方法, on the other hand, 通过聚类方法，快速定位出目标区域（焊缝区域），得到焊缝定位信息and finally,将G-CLAHE的特征增强与损失定位信息作为额外的两个通道与原始的单通道数据合并成一个三通道的图像数据。
aG-CLAHE method
图像增强处理技术It refers to highlight certain information in an image according to a particular need, at the same time, weaken or remove certain undesirable processing method of the information. The histogram statistics of the results, called by the image histogram equalization method 直方图均衡化. Histogram equalization can achieve the role of enhanced image display. The basic idea is the histogram of the original image is converted into the form of uniformly distributed, thus increasing the dynamic range of the pixel gray values to achieve the effect of enhancing the contrast of the entire image.
By this histogram, it can be observed, various image brightness uneven distribution proportion mostly try to increase the brightness in the histogram between the pixel and a high proportion of other accounts for a smaller percentage of the difference between the pixel you can improve the display of the image. Briefly, the method is enhanced histogram gradation range compression ratio less histogram pixels occupied, the extra space allocated to gradation histogram using a high proportion of the pixels to scale.
And G-CLAHE is increased stretched in the adaptive gradation CLAHE method, a method of improving CLAHE became G-CLAHE method. Gradation drawing method shown in FIG.
b. clustering methods
K-Means method used here, K-Means clustering algorithm firstly manually set according to the value of k, k cluster centers randomly chosen. Then calculates the distance between each object and the center of each subgroup, and assigned to the cluster center closest distance to each object. Assigned to each cluster center as its object a cluster. After completion of assignment, the center of each cluster is calculated based on the existing cluster of all the objects in the cluster. This process is repeated until the termination condition is satisfied.
2. Multi-scale discriminant model
Use a sliding window strategy combines deep learning model classifier constructed to locate damage. Using a sliding window slides on the raw data, determines whether each data covered by the sliding window comprises damage during sliding. The traditional sliding window strategy is difficult to ensure quality testing.
So here we are using multi-scale discriminant model. The model may utilize different scales classification determination area to be detected, to ensure that the neural network may get more information at the same space, with a rectangular frame positioning injury more granular.
使用S1 * S1大小的训练数据集对Location模型进行迁移学习的重新训练。然后利用S2 * S2像素大小对Deatil模块进行重新训练。将Location模块与Detail模块利用各自的数据分别训练过后，将两个模块以上图所示集成到多尺度模型框架中，进行联合训练。
3.主动学习方法
将多尺度多特征的损伤定位方法与主动学习的策略相结合，在Location模块的训练过程中不使用全部的训练数据，而是使用主动学习的方式，通过训练多个相同结构的分类器，以集成学习的方式组成一个更加强大的Location分类器模块。在不影响模型检测性能的前提下，尽可能的减少数据标注的相关成本。
Location 网络由 m 个子分类器 MC = (C1,C2, · · · ,Cm) 组成，其中每个子分类器 Ci 都可以独自完成 X 光图像的分类任务。当一个样本 I 输入到 Location 网络中之后，所有子分类器 MC = (C1,C2, · · · ,Cm) 均将根据自身的参数对输入数据进行计算并且得到相应的分类结果 MR = (R1, R2, · · · , Rm) ，由于本文中的任务是对每个样本进行有无损伤的二分类，所以其中每一个子分类计算结果 Ri 均是一个维度为1且长度为2的向 Ri = (R1i, R2i) ，其中 R1i 表示第 i 个分类器
认为该样本不包含损伤的置信度， R2i 表示第i个分类器认为该样本包含损伤的置信度。在训练过程中，将通过各个子分类器的分类结果 MR = (R1, R2, · · · , Rm) 以及样本选择策略来选择待标注的样本。在使用过程中，将通过各个子分类器的分类结果 MR = (R1, R2, · · · , Rm) 结合集成学习的投票策略来做出最终的分类决策。
基于目标检测改进的损失定位模型
结合目标识别的Yolo方法，对上述的模型做进一步的提升。
首先采用多特征融合方法将原始的单通道图像进行特征增强，得到多特征融合数据，然后根据滑动窗口的策略，利用步长为S/2,大小为 S * S 的窗口在多特征融合数据上进行采样，将采样结果及其在原始图像上的相对位置输入到LOcation模型中.Location模型通过集成学习的方式，对输入的S * S的图像是否包含损伤进行判别，如果其中包含损伤，则调用Detail模块中的YOLO网络对其中的损伤进行具体的定位。
