Finding a High Accuracy Neural Network for the Welding Defects Classification Using Efficient Neural Architecture Search via Parameter Sharing

Recent studies have shown that convolutional neural networks are achieving the best performance in image classification problems [1]. Thus, welding defects inspection in factory automation process can be performed using convolutional neural networks to determine whether welds on a mechanical part are defective or not. In deep learning area, it is well-known that finding the proper neural network architecture for specific task is highly difficult because there are so many available structures to choose from. Therefore, in this paper, we test and evaluate a method to select the novel convolutional neural network to determine whether the architecture search method is effective for the welding defect images. The method is based on using Efficient Neural Architecture Search via parameter sharing(ENAS) [2]. Using ENAS, we were able to find an architecture that achieved 0% error for 1,322 test images. Also, in the case of the MNIST dataset, we could find a novel architecture that achieved 99.77% accuracy for 10,000 test images.