Effective Feature Selection Technique for Deep Learning-Based Weld Defect Classification in Gamma Radiographic Images

Document Type : Original Article

Authors

1 Engineering Department, Nuclear Research Center, Egyptian Atomic Energy Authority (EAEA), Egypt.

2 Engineering Department, Nuclear Research Center, Egyptian Atomic Energy Authority (EAEA), Egypt

Abstract

The traditional defect detection techniques have poor detection accuracy and are strongly influenced by the industrial imaging environment. Concentrating on this shortcoming, this study presented a hybrid Deep Learning (DL) approach for automatic weld defects classification in Gamma Radiography Images (GRIs). We focus on improving accuracy by fusion and selection of deep-learned features extracted from five different DL models (e.g. SqueezeNet, GoogleNet, ShuffleNet, DarkNet19 and MobileNet-V2). To extract robust features from the DL models, the Pearson Correlation Coefficient (PCC), F-score (FS), and ReliefF (RF) feature selection algorithms are evaluated. The RF algorithm achieved the best result. The selected features are used for classification tasks using a Multiclass K-Nearest Neighbors (MKNN) classifier. Eight main types of weld defects and the normal type are considered in the utilized RGIs dataset. Several experiments are performed using the traditional feature extraction methods, DL methods, feature selection algorithms and the proposed one. Their results are compared to evaluate the performance of the presented system. Several RF feature subsets were tested, and the 450-feature subset with the best classification performance was found. The results confirmed that the suggested strategy performs better than all traditional and DL methods with an overall classification accuracy (CA) of 99.75%.

Keywords

Main Subjects

Files

Share

How to cite

Statistics