Objective

2

Traditional defect image recognition algorithms should manually build and select the main characteristics of targets

and the main defect features that are commonly used are as follows:(1) shape characteristics

including circumference

area

aspect ratio

and circularity; (2) gray scale features

including the mean of the gray scale and the variance of the gray level; and (3) characteristics adopted by the Hu invariant moment method to extract the mathematical features of translation

proportion

and rotation invariances and the suitable classifier selected to recognize the targets according to the selected features. The most frequently used classifiers include support vector machines

Adaboost

and artificial neural networks. These methods utilize several target features and classifiers to detect defects

which are usually designed manually. The identification process is time consuming and has certain limitations in application. Therefore

this study proposes a defect detection method that is based on faster regions with convolutional neural networks (faster R-CNN) features. The method uses convolution networks to automatically extract features of targets and avoids the problem of the dependence of defect detection on the character of the manual design.

Method

2

The method is based on a deep convolutional neural network. First

defect detection is determined. Three types of defects (detection targets) mainly exist in industrial computed tomography (CT) images

namely

slag inclusion

bubbles

and cracks. Then

the target images are manually annotated with a rectangular box (ground truth[GT] box)

the coordinate files for the GT are generated

and 42 types of anchor boxes are selected according to the aspect ratio of the GT bounding box. The Laplace operator and the homomorphic filter are used to sharpen and enhance the data set before training. The enhanced images are processed by the convolution and pooling layers. The convolution feature map is obtained and sent to the region proposal networks for the initial judgment of targets (no specific category) and background. Meanwhile

the target border is roughly regressed. Then

the generated proposal box is sent to the region of interest pooling layer to produce fixed-length output. Finally

the proposed feature maps are judged by the fully connected and softmax layers for specific categories

and the probability vector of each category is generated. Moreover

an accurate target detection box is regressed using the bounding box regression layer.

Result

2

The size of the picture in the data set to be detected is between 150×150 and 350×250

and each picture contains several different categories of bubbles

slags

and cracks. The trained model can be used to detect the defect images and effectively classify them into various categories of defect targets. The trained model according to the GT aspect ratio selected anchor that can detect the smallest area of the bubble is 9×9 piexl (that is

the size of the GT area

non-defect area)

and the smallest area with slag inclusion is 9×10 piexl. This model can accurately and quickly mark the locations of bubbles

slags

and cracks with a detection accuracy of up to 96%

and the average detection time of each image is 86 ms.

Conclusion

2

This work uses a deep convolutional neural network to automatically extract the features of defects and classify and identify defects on the basis of the extracted features

thereby avoiding the problem of manually selecting target features in traditional defect detection. Moreover

the process of defect recognition and localization becomes increasingly automated

thus resulting in reduced detection time and improved efficiency. The trained model according to the GT aspect ratio selected anchor is more pertinent for small targets

such as bubbles and slags

and for relatively large targets

such as cracks

and the border regression is accurate. The collected industrial CT images are grayscale pictures

whose information are not as rich as that in RGB pictures. Thus

CT image preprocessing before training can improve the detection accuracy of the training models

and the accuracy rate of the overall detection is better by 2% than that without preprocessing. The faster R-CNN defect detection algorithm proposed in this work has a fine detection effect and can detect the target in the industrial CT image quickly and accurately. If additional types of defects need to be detected

then a new detection model can be obtained by the fine-tuned training of the network. This work provides a new and efficient solution for the defect detection of industrial CT images.