Feature extraction from images and application of machi̇ne learning in the detecti̇on of casting defects

The detection of surface defects in casting processes is of significant importance for improving product quality and identifying production errors at early stages. In this study, an image processing and machine learning-based classification system has been developed to automatically detect defects from casting part images. The dataset used in the study was obtained from Kaggle and includes a total of 6,625 labeled images (3,750 defective and 2,875 non-defective). To enhance the classification capability of the images, a total of ten different feature extraction techniques were applied: Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), Scale-Invariant Feature Transform (SIFT), Gabor filters, Gray Level Co-occurrence Matrix (GLCM), wavelet transform, color histograms, Zernike moments, Hu moments, and Fourier transform. The effects of HOG and LBP features were evaluated using five different CellSize configurations (25x25, 50x50, 100x100, 150x150, and 300x300), and distinct feature datasets were generated for each variation. Due to the high dimensionality of the extracted feature vectors, feature selection algorithms such as minimum Redundancy Maximum Relevance (mRMR) and Chi-square were applied to identify the most informative features. The classification models were then retrained using these reduced datasets. Model training was performed using MATLAB's Classification Learner toolbox, and six different classification algorithms were tested and compared: Fine Tree, Fine K-Nearest Neighbor (KNN), Wide Neural Network, Bagged Trees, Gaussian SVM, and Binary Generalized Linear Model (GLM). According to the results, the highest classification accuracy of 99.7% was achieved with the Wide Neural Network model trained on the full feature set extracted at 25x25 CellSize. Feature selection with mRMR and Chi-square preserved model performance to a considerable extent while significantly reducing training time by lowering the feature dimensions. In the low-dimensional 300x300 CellSize configuration, feature selection was not applied, and high performance was still achieved with the original features. Overall, the proposed system demonstrated high accuracy and computational efficiency, indicating strong potential for integration into industrial casting quality control processes.