Surface Defect Detection using Hierarchical Features

Abstract

In this paper, we propose an instance level hierarchical features based convolution neural network model (H-CNN) for detecting surface defects. The H-CNN uses different convolutional layers’ extracted features to generate defect masks. The H-CNN first generates proposal regions. Then, it proposes a fully convolutional neural network to extract different level’s convolutional features and detect instance level defects. We applied the H-CNN model in freight train detection system for detecting oil-leaks, and the results demonstrate that the H-CNN can effectively identify and generate defect masks. It achieves 92% accuracy on the large reflective oil-leak stain, 86% on the large non-reflective oil-leak stain, 89% on the small reflective oil-leak stain and 74% on the small non-reflective oil-leak stain. Its image process speed is 0.467 s per frame.