A Novel Compact Convolutional Neural Network for Real-Time Nondestructive Evaluation of Metallic Surfaces

Abstract

Due to heightened concerns about quality assurance in the industry, the inspection of high-volume metallic parts has gained significant attention. In this article, we present a compact convolution neural network (CNN)-based optical nondestructive inspection scheme for high-volume metallic components. Our approach enables us to identify surface flaws as small as 2000th of an inch while discriminating them from flaw-like features, such as glue spots, dust, stains, and scratches. First, we capture the images of the metallic parts with a pixel size of $7000\times2500$ and label the images according to flawed and unflawed regions with the help of human experts. Second, we preprocess these images and form more than 120000 tiles of size $224\times 224\times3$ , which, subsequently, are fed into our compact CNN. Finally, we train the neural network by fine-tuning the hyperparameters to reduce overfitting of the data and to obtain maximum accuracy. The quantitative evaluation of this network shows that it can achieve almost 99% precision and recall of over 60 000 test images. The total image acquisition and inference time for one part-under-test is approximately 10 s. As a result, this novel compact CNN for optical inspection of surface flaws can outperform visual/manual inspection of high-volume metallic parts with very high accuracy and within a reasonable evaluation time.