Mitchins/anime-foreground-text-filter-v3

Anime/VN Overlay Text Filter (EfficientNet-B2, 640px)

Binary image classifier for filtering assets with overlaid communication text.

Labels:

• 0: no_text
• 1: has_text

Objective used during curation:

• has_text = 1 if the image contains text that is non-organic to the depicted scene and intentionally overlaid for communication.

Checkpoint

• Backbone: efficientnet_b2
• Input size: 640
• Format: safetensors
• File: model.safetensors
• Source run: runs/backbone_res_sweep_20260222_211117/tf_efficientnet_b2__img640
• Selection monitor during sweep: ood_fixed05_h_bal

Recommended Threshold

• 0.50 (balanced deployment default)

OOD test (anime_foreground_text_eval/ood_test_v2.csv) at thr=0.50:

• TP: 98
• FP: 6
• TN: 92
• FN: 4
• Precision: 0.9423
• Recall: 0.9608
• F1: 0.9515
• H_bal: 0.9497

Note: no threshold reached precision >= 0.98 on OOD dev for this checkpoint.

OOD Visual Audit (Recommended Threshold)

Green check = correct prediction, red X = incorrect prediction.
Images are linked to 4x retina versions.

Ground truth has_text:

Ground truth no_text:

Quick Inference

python inference.py --image path/to/image.jpg
python inference.py --glob "path/to/images/*.png" --threshold 0.50 --json

Notes

• Use threshold_tuning.json for full threshold sweep details.
• Re-tune threshold on your own live sample if deployment domain shifts.