Update README.md

README.md

@@ -1,3 +1,49 @@
 ---
 license: apache-2.0
+tags:
+  - vision
+  - image-classification
+  - dinov2
+  - moire-detection
+  - screen-detection
+  - anti-spoofing
+  - digital-forensics
+datasets:
+  - soumikrakshit/uhdm-dataset
+pipeline_tag: image-classification
 ---
+
+# Dual-Branch DINOv2 Moiré & Screen Recapture Detector
+
+A robust digital forensics classifier built on **DINOv2 (with registers)** designed to detect screen recaptures and Moiré patterns. 
+
+By combining an un-resized native crop (capturing high-frequency pixel interference) with a global thumbnail (capturing screen-wide periodic banding), this model overcomes the traditional scale dilemma in Moiré detection.
+
+---
+
+## 📁 Repository Files
+
+* **`best_screen_detector_backbone.pt`**: Weights for the fine-tuned top transformer blocks of the DINOv2 backbone.
+* **`best_screen_detector_mlp.pt`**: Weights for the 2-layer classification MLP head.
+* **`classes.json`**: Class index mapping (`0: "gt"`, `1: "moire"`).
+
+---
+
+## 🧠 Architecture Highlights
+
+1. **Dual-Branch Input**:
+   * **Local Branch (Native Resolution Crop):** Recropped directly at $224 \times 224$ from native image resolution without prior resizing, preventing antialiasing from washing out high-frequency Moiré lines.
+   * **Global Branch (Thumbnail View):** Resized to $256 \times 256$ and cropped to $224 \times 224$ to detect macroscopic periodic interference patterns.
+2. **Feature Concatenation ($3072$-d)**:
+   * Combines `[CLS_local, PatchMean_local, CLS_global, PatchMean_global]`, ignoring internal register tokens.
+3. **Partial Backbone Fine-Tuning**:
+   * Top 2 transformer layers adapted to Moiré patterns with differential learning rates, while earlier feature representations remain frozen.
+
+---
+
+## 🚀 Quick Start & Inference
+
+### 1. Requirements
+
+```bash
+pip install torch torchvision transformers pillow