Automatic defect detection from X-ray Scans for Aluminum Conductor Composite Core Wire Based on Classification Neutral Network

Abstract

The internal defects of the Aluminum Conductor Composite Core (ACCC) can be effectively visualized by X-ray imaging. However, the online detection of internal defects in the carbon core is challenging due to the huge data quantity and inconspicuous defects. In this paper, we propose an automatic defect detection framework for ACCC X-ray scans based on deep convolution neutral networks. The proposed method is based on an image classification network, with Inception-Resnet as backbone. Image normalization is conducted as the pre-processing step to reduce the training cost. We also propose a new data augmentation method according to the morphological character of the defect samples, to improve the extensiveness of the method. Experiment results have proven that the proposed method can effectively recognize the small and inconspicuous defects, with 3.5% improvement on mean Average Precision, compared to the object detection network (RetinaNet). Experiment also verifies the positive effect for data normalization and augmentation.