Improving Scattered Defect Grading in Castings Digital Radiographs via Smoothing the One-Hot Encoding

Abstract

Ensuring the precise grading of discontinuities is imperative to guarantee the quality of castings and enhance profitability in casting production. Recent grading methods leveraging computer vision are advanced by performing a single-label image classification or regression, which loses the intrinsically ordinal relationship. Motivated by this observation, we propose a label smoothing technology for ordinal variables to convert the level of each defect instance into a discrete probability distribution, aiming to model the noise label and ordinal relationship. Furthermore, we design a convolutional neural network framework based on multi-task learning. This framework, by simultaneously learning the level label distribution and regressing the level directly, outperforms a single-task network in terms of overall performance. Finally, we construct a casting gas porosity defect grading dataset. Experimental results on this dataset highlight the significant advantages of our proposed method compared to traditional single-label image classification or regression algorithms.