Update app.py
Browse files
app.py
CHANGED
|
@@ -161,10 +161,12 @@ def segment_lines_yolo(bin_arr, yolo_model):
|
|
| 161 |
|
| 162 |
_surya_predictor = None
|
| 163 |
|
| 164 |
-
def segment_lines_surya(
|
| 165 |
global _surya_predictor
|
| 166 |
-
|
| 167 |
-
|
|
|
|
|
|
|
| 168 |
|
| 169 |
try:
|
| 170 |
from surya.detection import DetectionPredictor
|
|
@@ -174,16 +176,18 @@ def segment_lines_surya(bin_arr):
|
|
| 174 |
if _surya_predictor is None:
|
| 175 |
_surya_predictor = DetectionPredictor()
|
| 176 |
|
| 177 |
-
|
| 178 |
-
|
|
|
|
|
|
|
| 179 |
predictions = _surya_predictor([pil_image])
|
| 180 |
|
| 181 |
-
CONF_THRESHOLD = 0.6
|
| 182 |
raw = []
|
| 183 |
if predictions and predictions[0].bboxes:
|
| 184 |
for bbox in predictions[0].bboxes:
|
| 185 |
conf = getattr(bbox, "confidence", 1.0)
|
| 186 |
-
if conf < CONF_THRESHOLD:
|
| 187 |
continue
|
| 188 |
x1, y1, x2, y2 = bbox.bbox
|
| 189 |
if (x2 - x1) > 5 and (y2 - y1) > 5:
|
|
@@ -192,6 +196,7 @@ def segment_lines_surya(bin_arr):
|
|
| 192 |
raw.sort(key=lambda b: (b[1] + b[3]) / 2)
|
| 193 |
|
| 194 |
def overlaps_v(a, b):
|
|
|
|
| 195 |
return a[1] < b[3] and b[1] < a[3]
|
| 196 |
|
| 197 |
merged = []
|
|
@@ -205,9 +210,9 @@ def segment_lines_surya(bin_arr):
|
|
| 205 |
if not placed:
|
| 206 |
merged.append(list(box))
|
| 207 |
|
|
|
|
| 208 |
merged.sort(key=lambda b: b[1])
|
| 209 |
return [(int(b[0]), int(b[1]), int(b[2]-b[0]), int(b[3]-b[1])) for b in merged]
|
| 210 |
-
|
| 211 |
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 212 |
# 3. WORD SEGMENTATION
|
| 213 |
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
|
|
|
| 161 |
|
| 162 |
_surya_predictor = None
|
| 163 |
|
| 164 |
+
def segment_lines_surya(pil_image):
|
| 165 |
global _surya_predictor
|
| 166 |
+
|
| 167 |
+
# 1. Ensure the image is exactly RGB as Surya expects (like your local script)
|
| 168 |
+
if pil_image.mode != "RGB":
|
| 169 |
+
pil_image = pil_image.convert("RGB")
|
| 170 |
|
| 171 |
try:
|
| 172 |
from surya.detection import DetectionPredictor
|
|
|
|
| 176 |
if _surya_predictor is None:
|
| 177 |
_surya_predictor = DetectionPredictor()
|
| 178 |
|
| 179 |
+
# Set thresholds globally right before inference (replicating your local os.environ setup)
|
| 180 |
+
os.environ["DETECTOR_TEXT_THRESHOLD"] = "0.75"
|
| 181 |
+
os.environ["DETECTOR_BLANK_THRESHOLD"] = "0.45"
|
| 182 |
+
|
| 183 |
predictions = _surya_predictor([pil_image])
|
| 184 |
|
| 185 |
+
CONF_THRESHOLD = 0.6 # Matching your local script
|
| 186 |
raw = []
|
| 187 |
if predictions and predictions[0].bboxes:
|
| 188 |
for bbox in predictions[0].bboxes:
|
| 189 |
conf = getattr(bbox, "confidence", 1.0)
|
| 190 |
+
if conf is not None and conf < CONF_THRESHOLD:
|
| 191 |
continue
|
| 192 |
x1, y1, x2, y2 = bbox.bbox
|
| 193 |
if (x2 - x1) > 5 and (y2 - y1) > 5:
|
|
|
|
| 196 |
raw.sort(key=lambda b: (b[1] + b[3]) / 2)
|
| 197 |
|
| 198 |
def overlaps_v(a, b):
|
| 199 |
+
# Your local script uses this exact simple overlap logic
|
| 200 |
return a[1] < b[3] and b[1] < a[3]
|
| 201 |
|
| 202 |
merged = []
|
|
|
|
| 210 |
if not placed:
|
| 211 |
merged.append(list(box))
|
| 212 |
|
| 213 |
+
# Sort top-to-bottom, output as [x, y, w, h] (Matching your local script output format)
|
| 214 |
merged.sort(key=lambda b: b[1])
|
| 215 |
return [(int(b[0]), int(b[1]), int(b[2]-b[0]), int(b[3]-b[1])) for b in merged]
|
|
|
|
| 216 |
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 217 |
# 3. WORD SEGMENTATION
|
| 218 |
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|