[ad_1]
Context: Morphological image analysis techniques are essential for extracting meaningful features from complex image data, especially in pattern recognition tasks like handwritten digit classification.
Problem: Traditional feature sets often struggle to balance global and local image structures, limiting classification accuracy and model interpretability, particularly on small datasets.
Approach: We propose a hybrid pipeline that fuses Adaptive Morphological Reconstruction (AMR) features with Haralick texture descriptors and raw pixel values. A random forest classifier is trained using this enriched feature set, with hyperparameter tuning and cross-validation to optimize performance.
Results: The combined approach achieves a test accuracy of 97.2% on the scikit-learn digits dataset. Feature importance analysis shows that while raw pixels are dominant, texture and AMR features also provide valuable discriminatory information.
Conclusions: Integrating classical feature engineering with raw data significantly enhances classification accuracy and model robustness, illustrating the ongoing value of adaptive morphological methods in modern image analysis workflows.
Keywords: adaptive morphological reconstruction; image classification; feature engineering; random forest; digit recognition
[ad_2]
