Denoising Autoencoder and Histogram-GLCM Feature Fusion for Image-Based Classification of Locally Made Earthen Ceramic Pots

This study presents an image-based classification approach for assessing the quality of locally made earthen ceramic pots by integrating deep learning with handcrafted texture analysis. A Denoising Autoencoder with Convolutional Layers (DAE-CAE) was employed to reduce image noise, including Gaussian blur and salt-and-pepper distortions. After denoising, Histogram Equalization and Gray-Level Cooccurrence Matrix (GLCM) techniques were applied to enhance contrast and extract texture features. Using a dataset of 1,007 labeled images expanded to 2,768 via data augmentation, the fused features were used to train classifiers including Support Vector Machine (SVM), Random Forest, and CNN. The SVM model with the combined feature set (Model A3) achieved the best performance, with 97.1% accuracy and a ROC AUC of 0.996, outperforming models using individual features. These results demonstrate the effectiveness of combining intensity-based and texture-based features for improved classification accuracy, generalization, and scalability. The proposed method is suitable for smallscale manufacturing and adaptable to other surface quality inspection tasks. Future work includes integrating attention mechanisms, transformer-based models, and expanding to real-time, multi-class quality grading systems.