Convolution With Rotation Invariance for Online Detection of Tiny Defects on Magnetic Tile Surface

Automated surface defect detection for products attracts lots of interest in the industry, which is typically an image classification task. Although convolutional neural networks (CNNs) with shift and scale invariance have become the de-facto standard for such task in computer vision, their applications to millimeter-level tiny defects with random orientation on the dark surface with the complex texture of magnetic tile remains limited. In this paper, we propose, for the first time, introducing rotation invariance into convolution by rotating feature maps, thus enabling CNNs to actively detect defects from multiple directions and decreasing the probability of missing tiny defects. Next, we establish a lightweight neural architecture, i.e., CNN with rotation invariance (RICNN), with a model size of only 0.119M. A developed image collection system can obtain real-time images on magnetic tile surface. Additionally, we train and evaluate the model performance on a built MTSD3C6K dataset and a publicly available IACAS3C6K dataset. Experimental results on both datasets show that the proposed RICNN obtains superior classification performance compared with state-of-the-art networks while requiring substantially fewer computational costs. The source code is available at https://github.com/Clarkxielf/Convolution-with-Rotation-Invariance-for-Online-Detection-of-Tiny-Defects-on-Magnetic-Tile-Surface.