Publication: Detection of void of single pad in x-ray image using deep convolutional neural network
Loading...
Date
2020-07-01
Authors
Mohd Azran, Muhammad Amin Khalis
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Early studies on deep convolutional neural network (DCNN) reveal a promising method in order to detect and segment the desired areas as CNN able to capture the spatial
and temporal dependencies in an image. Motivated by this finding, this research proposes to develop voids detection system from x-ray images using DCNN. Manual detection of
void of single pad from x-ray images is time consuming and too subjective as the detection performance depends on human skills and knowledge. Thus, automated defect
detector could ease the inspection and assist the workers in order to improve the detection performance of voids or defects. To the best of my knowledge, there still no previous
research that proposed a DCNN system in order to detect the void of single pad from x ray images. Three datasets of x-ray images namely Kb-data, SC-data and Bb-data are
obtained from ViTrox Technologies Sdn. Bhd. In the proposed system, those images are first augmented through rotating and mirroring processes to obtain more images in order
to assist the training of DCNN for better detection performance. Then, the original x-rays images and their ground truth images (showing the locations of voids) are fed into U-Net
DCNN for segmentation process of void areas. The results show that the developed system has successfully produced accuracy of 94.57%, 97.53% and 98.58% for Kb-data,
SC-data and Bb-data datasets respectively. Furthermore, the average computational time for voids areas segmentation for a single X-ray image is 1.5 seconds, 93.75 milliseconds
and 0.15 seconds respectively shorter than manual detection which is takes 6 seconds in average. The advantages of the proposed method are the fast-computational time for voids
areas segmentation, does not need extra method to improve its performance and easy to calibrate if compared to previous methods. Therefore, the proposed method could possibly provide a reliable method for the development of automated voids detection system which will aid the workers in manufacturing industry.