AsTexNet: A Discriminative Scale-Aware Hybrid Texture Representation with Compact Embedding

Texture classification remains challenging due to high inter-class similarity, scale variability, and limited labeled data. While handcrafted descriptors capture fine-grained microstructures and CNNs encode global semantics, existing hybrid approaches often rely on direct feature concatenation or heuristic fusion strategies that ignore the varying discriminative importance of texture scales. Moreover, many deep texture representations introduce high-dimensional redundant embeddings and require computationally expensive backbone adaptation, limiting their effectiveness in small-data scenarios. This paper introduces AsTexNet, a discriminative scale-aware hybrid representation framework that reformulates multi-scale feature integration as a structured weighting problem. The proposed method quantifies the class-separability contribution of each handcrafted scale using a data-driven weighting mechanism, enabling adaptive emphasis on informative texture patterns. In parallel, pretrained CNN features are enhanced using Generalized Mean (GeM) pooling and compacted via cross-validated PCA to improve separability while controlling redundancy. A normalized late-fusion formulation integrates local and global cues into a compact 512-dimensional embedding without requiring backbone fine-tuning. Extensive experiments on six benchmark datasets demonstrate consistent performance gains, achieving 100% accuracy on Kylberg and Brodatz, 99.4% on UIUC, 99.0% on UMD, 90.6% on KTH-TIPS2b, and 78.71% on FMD. These results demonstrate that explicitly modeling scale-wise discriminative contributions leads to more effective and compact texture representations, particularly in limited-data scenarios.