Genty: Multi-Scale Character, Word & Sentence Detection CNN (Fine-Tuned)
This repository contains the fine-tuned checkpoint of TextDetectionCNN, optimized for high-precision text localization.
π Model Architecture & Multi-Scale Heads
| Scale | Feature Grid | Stride | Anchor Detection Target |
|---|---|---|---|
| π΅ Character | $64 imes 64$ | 8 | Individual characters & symbols |
| π’ Word | $32 imes 32$ | 16 | Tokens, words & numeric strings |
| π΄ Sentence | $16 imes 16$ | 32 | Complete sentences & paragraphs |
π Fine-Tuning Performance & Results
- Best Validation mAP@0.50:
0.3333(Achieved at Epoch 4) - Input Resolution:
512 x 512(Aspect-Ratio Preserving Letterbox) - Loss Function: Sigmoid Focal Loss (Objectness) + Smooth L1 (Coordinates)
- Hardware: 2x NVIDIA Tesla T4 GPUs (AMP FP16 Acceleration)
- Downloads last month
- 89