Deep Residual Neural Network-Based Defect Detection on Complex Backgrounds

• Published in: IEEE transactions on instrumentation and measurement Vol. 71; pp. 1 - 10
• Main Authors: Ho, Chao-Ching, Benalcazar Hernandez, Miguel A., Chen, Yi-Fan, Lin, Chih-Jer, Chen, Chin-Sheng
• Format: Journal Article
• Language: English
• Published: New York IEEE 2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Annotations; Artificial neural networks; Classification; Datasets; Deep learning; deep residual neural network (DRNN); defect detection; Defects; Feature extraction; Image acquisition; Image segmentation; Inspection; Neural networks; Plastics; Residual neural networks; ResnNet50; Robot arms; Training
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2022.3144224

An inspection device with a set of lights, a six-axis robot arm, and a camera is designed for image acquisition. A deep residual neural network (DRNN) performs both the feature extraction and classification tasks simultaneously. This article makes several modifications of ResnNet50 to ensure accurate and reliable predictions. The dataset consists of a collection of images of a plastic casing with different types of scratches. Most defects in the dataset are thin, shallow, and small. Due to the spatial reduction, defects fade because of their features while performing the deep residual network. Thus, mask labeling based on pixel annotation crops a subimage using a sliding window. Efforts have been made to solve the faded issue while DRNN. The experimental results for 600 plastic casing images show that the proposed method significantly increases the convolutional algorithm's ability to detect defects accurately. The defect detection accuracy is approximately 96.38%.