A hybrid deep learning framework with fractional multi-scale directional texture features for automated waste classification

Accurate waste image classification is essential for intelligent recycling and sustainable waste management systems. However, reliable waste recognition remains challenging due to irregular object shapes, illumination variations, complex backgrounds, and limited annotated data. Although deep learning methods have achieved promising results, they often require large training datasets and may not effectively capture fine-grained texture characteristics necessary for distinguishing visually similar waste categories. This study proposes a hybrid handcrafted–deep learning framework for automated waste classification. The approach employs a Fractional Multi-Scale Directional Texture Descriptor (FMSDTD) to extract discriminative fractional-order, multi-scale, and directional texture features from waste images. These handcrafted features effectively characterize material-specific surface patterns commonly found in plastics, metals, paper, cardboard, glass, and organic waste. The extracted texture features are fused with deep features generated by a lightweight Convolutional Neural Network (CNN), enabling complementary representation of local textures and high-level shape information while maintaining computational efficiency. The combined feature set is subsequently classified using a Support Vector Machine (SVM) for nine-class waste recognition. Experiments conducted on the publicly available RealWaste dataset demonstrate the effectiveness of the proposed framework, achieving a classification accuracy of 97.32%. The results indicate that integrating FMSDTD with CNN features enhances feature discriminability and robustness, providing an accurate and computationally efficient solution for intelligent waste sorting and recycling applications.