NDTNet: Optical Nondestructive Evaluation with Compact Convolutional Neural Network

摘要

The inspection of high volume parts is critical for quality assurance in manufacturing industries. In this article, we put forth NDTNet, a compact convolution neural network (CCNN) based surface flaw inspection of metallic components. The compactness of CNN is achieved by model compression techniques like parameter pruning and quantization, which reduces its size by a factor of 18. In order to account for loss due to model compression, the CNN is trained with iterative parameter pruning and simulated quantization. The compact CNN is trained on over 120,000 images of size 224-by-224-by-3. The statistical evaluation of the proposed compact CNN shows it can achieve accuracy, precision and recall in the range of 96-98% for different compression techniques. The total evaluation time i.e. image acquisition and inference is about 10 seconds. Therefore, the proposed compact CNN is suitable for high volume metallic component inspection in real-time with high accuracy and speed.