A Lightweight Transformer for Non-Invasive Ivory Texture Recognition

Non-invasive ivory identification is vital for wildlife conservation and trade regulation, but current techniques remain invasive, slow, or unsuitable for real-time use. Therefore, we propose LT-Former, a lightweight vision transformer tailored for fine-grained ivory texture classification. Specifically, we introduce a portable, non-invasive texture-capture system with automated patch selection; next, we integrate efficient channel attention, mixed local channel attention, and Linformer self-attention to recalibrate, enhance, and globally mix features; meanwhile, we build IvoryTex, a high-resolution benchmark of labeled texture patches. Further, under identical training conditions, LT-Former outperforms state-of-the-art lightweight vision transformers across multiple patch scales, achieving 94.15% Top-1 accuracy on 256×256 patches with 11.3M parameters, 2.2G FLOPs, and 1.70 ms latency, enabling real-time, non-invasive ivory recognition.