2019 Volume 22 Issue 6 Pages 559-567
This paper considers augmenting training images composed of defect and background images in the training phase when a Region-Based Convolutional Neural Network (R-CNN) is applied to an appearance inspection. This approach is applied to the appearance inspection of plastic pieces produced by a press work and the obtained results are reported. In the proposed image composition, firstly, some typical defective patterns are cut from the actual images and then, these defective patterns are pasted to the actual background images by changing the size, rotation, colors, converting them to black-and-white, etc. In the experiments, 81 defective patterns are cut from 50 actual images and pasted on several background images producing 500 unique images by applying the image modifications. R-CNN is trained on these 500 images and then applied to the detection of 53 defective patterns in 40 images. The resulting detection ratio and hit ratio are 81% and 86%, respectively indicating that the suggested approach is promising for practical use with some performance improvements.
