Spaces:
Running on Zero
Running on Zero
File size: 7,139 Bytes
99857c5 694bc64 99857c5 694bc64 99857c5 6e08ed4 99857c5 694bc64 99857c5 6e08ed4 99857c5 694bc64 99857c5 6e08ed4 99857c5 6e08ed4 99857c5 6e08ed4 99857c5 6e08ed4 99857c5 694bc64 99857c5 6e08ed4 99857c5 6e08ed4 99857c5 694bc64 99857c5 6e08ed4 99857c5 694bc64 99857c5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 | """Braille Reader — Upload a braille image, get English text."""
import json
import os
import tempfile
import uuid
from datetime import datetime
from pathlib import Path
import cv2
import gradio as gr
import numpy as np
import spaces
import torch
from huggingface_hub import CommitScheduler, hf_hub_download
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from ultralytics import YOLO
# --- Model loading (on CPU at startup, GPU allocated per-request) ---
YOLO_REPO = "prasanthmj/yolov8-braille"
BYT5_REPO = "prasanthmj/braille-byt5-v3"
DATASET_REPO = "prasanthmj/braille-reader-results"
print("Loading models...")
# YOLOv8 braille detector
weights_path = hf_hub_download(YOLO_REPO, "yolov8_braille.pt")
braille_map_path = hf_hub_download(YOLO_REPO, "braille_map.json")
yolo_model = YOLO(weights_path)
with open(braille_map_path) as f:
dot_to_unicode = json.load(f)
# ByT5 Grade 2 interpreter (load on CPU, moved to GPU per-request)
tokenizer = AutoTokenizer.from_pretrained(BYT5_REPO)
byt5_model = AutoModelForSeq2SeqLM.from_pretrained(BYT5_REPO)
byt5_model.eval()
print("Models loaded (CPU). GPU allocated per-request via ZeroGPU.")
# --- Result saving via CommitScheduler ---
RESULTS_DIR = Path("./results")
RESULTS_DIR.mkdir(exist_ok=True)
(RESULTS_DIR / "images").mkdir(exist_ok=True)
scheduler = CommitScheduler(
repo_id=DATASET_REPO,
repo_type="dataset",
folder_path=RESULTS_DIR,
every=5, # push every 5 minutes
token=os.environ.get("HF_TOKEN"),
)
def save_result(image: np.ndarray, braille_text: str, english_text: str,
total_cells: int, num_lines: int, avg_conf: float):
"""Save image and result to the dataset (batched by CommitScheduler)."""
entry_id = datetime.utcnow().strftime("%Y%m%d_%H%M%S") + "_" + uuid.uuid4().hex[:6]
# Save image
image_filename = f"images/{entry_id}.jpg"
image_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
cv2.imwrite(str(RESULTS_DIR / image_filename), image_bgr)
# Append to JSONL
record = {
"id": entry_id,
"image": image_filename,
"braille_unicode": braille_text,
"english": english_text,
"cells": total_cells,
"lines": num_lines,
"avg_confidence": round(avg_conf, 4),
"timestamp": datetime.utcnow().isoformat(),
}
with scheduler.lock:
with open(RESULTS_DIR / "results.jsonl", "a") as f:
f.write(json.dumps(record) + "\n")
# --- CLAHE Preprocessing ---
def preprocess_clahe(image_path: str) -> str:
"""Apply CLAHE preprocessing for better detection on low-contrast images."""
img = cv2.imread(image_path)
if img is None:
return image_path
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
enhanced = clahe.apply(gray)
enhanced_bgr = cv2.cvtColor(enhanced, cv2.COLOR_GRAY2BGR)
tmp = tempfile.NamedTemporaryFile(suffix=".jpg", delete=False)
cv2.imwrite(tmp.name, enhanced_bgr)
return tmp.name
# --- Stage 1: YOLOv8 Detection ---
def detect_braille(image_path: str, confidence: float = 0.15) -> list[list[dict]]:
"""Detect braille cells and group into lines."""
results = yolo_model.predict(image_path, conf=confidence, verbose=False)
boxes = results[0].boxes
if len(boxes) == 0:
return []
n = len(boxes)
data = np.zeros((n, 6))
data[:, 0] = boxes.xywh[:, 0].cpu().numpy()
data[:, 1] = boxes.xywh[:, 1].cpu().numpy()
data[:, 2] = boxes.xywh[:, 2].cpu().numpy()
data[:, 3] = boxes.xywh[:, 3].cpu().numpy()
data[:, 4] = boxes.conf.cpu().numpy()
data[:, 5] = boxes.cls.cpu().numpy()
# Sort by Y
data = data[data[:, 1].argsort()]
# Split into lines by Y gaps
avg_height = np.mean(data[:, 3])
y_threshold = avg_height / 2
y_diffs = np.diff(data[:, 1])
break_indices = np.where(y_diffs > y_threshold)[0]
raw_lines = np.split(data, break_indices + 1)
lines = []
for raw_line in raw_lines:
raw_line = raw_line[raw_line[:, 0].argsort()]
cells = []
for row in raw_line:
class_idx = int(row[5])
dots = yolo_model.names[class_idx]
unicode_char = dot_to_unicode.get(dots, "?")
cells.append({
"dots": dots,
"unicode": unicode_char,
"confidence": row[4],
})
lines.append(cells)
return lines
# --- Main pipeline (GPU allocated here) ---
@spaces.GPU
def transcribe(image) -> str:
"""Full pipeline: image -> detection -> interpretation -> English text."""
if image is None:
return "Please upload an image."
# Save uploaded image to temp file
tmp = tempfile.NamedTemporaryFile(suffix=".jpg", delete=False)
if isinstance(image, np.ndarray):
cv2.imwrite(tmp.name, cv2.cvtColor(image, cv2.COLOR_RGB2BGR))
else:
cv2.imwrite(tmp.name, image)
image_path = tmp.name
# CLAHE preprocessing
processed_path = preprocess_clahe(image_path)
# Stage 1: Detect braille cells
lines = detect_braille(processed_path)
if not lines:
return "No braille cells detected. Try a clearer image."
# Extract Unicode braille per line
braille_lines = ["".join(cell["unicode"] for cell in line) for line in lines]
# Stats
total_cells = sum(len(line) for line in lines)
avg_conf = float(np.mean([cell["confidence"] for line in lines for cell in line]))
# Stage 2: Interpret each line with ByT5 on GPU
device = "cuda" if torch.cuda.is_available() else "cpu"
byt5_model.to(device)
english_lines = []
for line in braille_lines:
if not line.strip():
english_lines.append("")
continue
input_text = f"translate Braille to English: {line}"
inputs = tokenizer(input_text, return_tensors="pt", max_length=1024, truncation=True)
inputs = {k: v.to(device) for k, v in inputs.items()}
with torch.no_grad():
outputs = byt5_model.generate(**inputs, max_length=512)
decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
english_lines.append(decoded)
# Format output
braille_text = "\n".join(braille_lines)
english_text = "\n".join(english_lines)
# Save to dataset
save_result(image, braille_text, english_text, total_cells, len(lines), avg_conf)
output = f"{english_text}\n\n"
output += f"--- Details ---\n"
output += f"Cells detected: {total_cells}\n"
output += f"Lines: {len(lines)}\n"
output += f"Avg confidence: {avg_conf:.1%}\n"
output += f"\nBraille Unicode:\n{braille_text}"
return output
# --- Gradio UI ---
demo = gr.Interface(
fn=transcribe,
inputs=gr.Image(type="numpy", label="Upload Braille Image"),
outputs=gr.Textbox(label="English Translation", lines=15),
title="Braille Reader",
description="Upload a scanned braille document to get its English translation. Supports Grade 2 (contracted) braille.",
examples=[],
flagging_mode="never",
)
if __name__ == "__main__":
demo.launch()
|