Nisha56 commited on
Commit
5ff8f69
Β·
1 Parent(s): 6bccba5

resolved feature issue, added library page

Browse files
Files changed (6) hide show
  1. crud.py +12 -1
  2. grounding.py +11 -37
  3. layer_extractor.py +11 -1
  4. main.py +48 -10
  5. ocr.py +1 -1
  6. proposal_engine.py +63 -9
crud.py CHANGED
@@ -69,4 +69,15 @@ def get_saved_state(db, file_id: str):
69
  "filename": project.filename,
70
  "file_id": project.file_id,
71
  "layer_count": project.layer_count,
72
- }
 
 
 
 
 
 
 
 
 
 
 
 
69
  "filename": project.filename,
70
  "file_id": project.file_id,
71
  "layer_count": project.layer_count,
72
+ }
73
+
74
+ def delete_project(db: Session, file_id: str):
75
+ project = db.query(Project).filter(Project.file_id == file_id).first()
76
+
77
+ if not project:
78
+ return None
79
+
80
+ db.delete(project)
81
+ db.commit()
82
+
83
+ return project
grounding.py CHANGED
@@ -4,10 +4,8 @@ import torch
4
  import numpy as np
5
  from PIL import Image
6
 
7
- # tell Python where groundingdino package is
8
  from groundingdino.util.inference import load_model, load_image, predict
9
 
10
- # config
11
  GROUNDING_CONFIG = os.path.join(os.path.dirname(__file__), "models", "GroundingDINO_SwinT_OGC.py")
12
  GROUNDING_WEIGHTS = os.path.join(os.path.dirname(__file__), "models", "groundingdino_swint_ogc.pth")
13
  DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
@@ -15,6 +13,7 @@ DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
15
  # detection thresholds
16
  BOX_THRESHOLD = 0.30 # confidence needed to keep a box
17
  TEXT_THRESHOLD = 0.25 # confidence needed to keep a text label
 
18
 
19
  print(f"[GDINO] Using device: {DEVICE}")
20
 
@@ -28,16 +27,8 @@ def load_grounding_model():
28
 
29
  def detect_objects(image_path: str, model, text_prompt: str = None) -> list:
30
  """
31
- Detect objects in image using text prompt.
32
-
33
- text_prompt examples:
34
- "product . text . logo . background"
35
- "person . car . sky"
36
- If None, uses a default poster-focused prompt.
37
-
38
- Returns list of dicts:
39
- [{label, confidence, x1, y1, x2, y2, cx, cy, w, h}]
40
- All coordinates are absolute pixels.
41
  """
42
  if text_prompt is None:
43
  # text_prompt = (
@@ -45,16 +36,14 @@ def detect_objects(image_path: str, model, text_prompt: str = None) -> list:
45
  # "image . graphic . illustration . icon . button"
46
  # )
47
  text_prompt = (
48
- "product . logo . person . icon . illustration . graphic"
49
  )
50
 
51
- # load_image returns (PIL image, transformed tensor) β€” both needed
52
  image_pil, image_tensor = load_image(image_path)
53
 
54
 
55
  img_h, img_w = image_pil.shape[:2]
56
 
57
- # run detection
58
  with torch.no_grad():
59
  boxes, confidences, labels = predict(
60
  model = model,
@@ -94,21 +83,17 @@ def detect_objects(image_path: str, model, text_prompt: str = None) -> list:
94
 
95
  # sort by confidence descending
96
  results.sort(key=lambda r: r["confidence"], reverse=True)
 
 
 
 
 
97
  return results
98
 
99
 
100
  def boxes_to_sam_prompts(detections: list) -> tuple:
101
  """
102
- Convert GroundingDINO detections into SAM2 input format.
103
-
104
- SAM2 accepts:
105
- - point_coords: array of (x, y) centre points
106
- - point_labels: array of 1s (foreground)
107
-
108
- We use the centre of each detected box as a SAM2 prompt point.
109
- This tells SAM2 exactly where to segment instead of using blind grid.
110
-
111
- Returns: (point_coords np.array, point_labels np.array, labels list)
112
  """
113
  if not detections:
114
  return None, None, []
@@ -128,8 +113,7 @@ def boxes_to_sam_prompts(detections: list) -> tuple:
128
 
129
  def save_detection_debug(image_path: str, detections: list, output_dir: str = None):
130
  """
131
- Save debug image showing GroundingDINO bounding boxes.
132
- Useful to visually verify detection quality before passing to SAM2.
133
  """
134
  import cv2
135
 
@@ -159,7 +143,6 @@ def save_detection_debug(image_path: str, detections: list, output_dir: str = No
159
  return out_path
160
 
161
 
162
- # run directly to test
163
  if __name__ == "__main__":
164
  BACKEND_DIR = os.path.dirname(os.path.abspath(__file__))
165
  TEST_IMAGE = os.path.join(BACKEND_DIR, "test_images", "sale_img.jpg")
@@ -177,15 +160,6 @@ if __name__ == "__main__":
177
 
178
  point_coords, point_labels, sam_labels = boxes_to_sam_prompts(detections)
179
 
180
- # print("\nSAM2 Prompt Points:")
181
- # print(point_coords)
182
-
183
- # print("\nSAM2 Point Labels:")
184
- # print(point_labels)
185
-
186
- # print("\nDetection Labels:")
187
- # print(sam_labels)
188
-
189
  print(f"\nDetected {len(detections)} objects:")
190
  for d in detections:
191
  print(f" {d['label']:20s} conf={d['confidence']} box=({d['x1']},{d['y1']}) β†’ ({d['x2']},{d['y2']})")
 
4
  import numpy as np
5
  from PIL import Image
6
 
 
7
  from groundingdino.util.inference import load_model, load_image, predict
8
 
 
9
  GROUNDING_CONFIG = os.path.join(os.path.dirname(__file__), "models", "GroundingDINO_SwinT_OGC.py")
10
  GROUNDING_WEIGHTS = os.path.join(os.path.dirname(__file__), "models", "groundingdino_swint_ogc.pth")
11
  DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 
13
  # detection thresholds
14
  BOX_THRESHOLD = 0.30 # confidence needed to keep a box
15
  TEXT_THRESHOLD = 0.25 # confidence needed to keep a text label
16
+ MIN_BOX_SIZE = 10
17
 
18
  print(f"[GDINO] Using device: {DEVICE}")
19
 
 
27
 
28
  def detect_objects(image_path: str, model, text_prompt: str = None) -> list:
29
  """
30
+ Detect objects using GroundingDINO.
31
+ Returns bounding boxes in image coordinates.
 
 
 
 
 
 
 
 
32
  """
33
  if text_prompt is None:
34
  # text_prompt = (
 
36
  # "image . graphic . illustration . icon . button"
37
  # )
38
  text_prompt = (
39
+ "product . logo . person . icon . illustration . graphic . shape . cup"
40
  )
41
 
 
42
  image_pil, image_tensor = load_image(image_path)
43
 
44
 
45
  img_h, img_w = image_pil.shape[:2]
46
 
 
47
  with torch.no_grad():
48
  boxes, confidences, labels = predict(
49
  model = model,
 
83
 
84
  # sort by confidence descending
85
  results.sort(key=lambda r: r["confidence"], reverse=True)
86
+
87
+ results = [
88
+ r for r in results
89
+ if r["w"] >= MIN_BOX_SIZE and r["h"] >= MIN_BOX_SIZE
90
+ ]
91
  return results
92
 
93
 
94
  def boxes_to_sam_prompts(detections: list) -> tuple:
95
  """
96
+ Convert detections into SAM2 point prompts.
 
 
 
 
 
 
 
 
 
97
  """
98
  if not detections:
99
  return None, None, []
 
113
 
114
  def save_detection_debug(image_path: str, detections: list, output_dir: str = None):
115
  """
116
+ Save detection boxes for debugging.
 
117
  """
118
  import cv2
119
 
 
143
  return out_path
144
 
145
 
 
146
  if __name__ == "__main__":
147
  BACKEND_DIR = os.path.dirname(os.path.abspath(__file__))
148
  TEST_IMAGE = os.path.join(BACKEND_DIR, "test_images", "sale_img.jpg")
 
160
 
161
  point_coords, point_labels, sam_labels = boxes_to_sam_prompts(detections)
162
 
 
 
 
 
 
 
 
 
 
163
  print(f"\nDetected {len(detections)} objects:")
164
  for d in detections:
165
  print(f" {d['label']:20s} conf={d['confidence']} box=({d['x1']},{d['y1']}) β†’ ({d['x2']},{d['y2']})")
layer_extractor.py CHANGED
@@ -75,6 +75,10 @@ def build_layers(image_path: str) -> tuple:
75
  print("[LAYERS] Running OCR...")
76
  text_blocks = extract_text_layers(image_path)
77
 
 
 
 
 
78
  # --- classify and route ---
79
  image_type, logo_detection = classify_image_type(detections, img_w, img_h)
80
 
@@ -120,7 +124,12 @@ def build_layers(image_path: str) -> tuple:
120
  return layers, groups
121
 
122
  # --- not-logo path ---
123
- clean_objects = run_proposal_engine(object_proposals, img_w, img_h)
 
 
 
 
 
124
  print(f"[LAYERS] {len(clean_objects)} objects passed proposal engine")
125
 
126
  layers = []
@@ -194,6 +203,7 @@ def build_layers(image_path: str) -> tuple:
194
  ]
195
 
196
  clean_text, rejected_text = reject_text_inside_objects(text_proposals, object_boxes)
 
197
  if rejected_text:
198
  print(f"[LAYERS] Rejected {len(rejected_text)} text blocks inside object layers")
199
 
 
75
  print("[LAYERS] Running OCR...")
76
  text_blocks = extract_text_layers(image_path)
77
 
78
+ print("\nOCR DETECTED:")
79
+ for t in text_blocks:
80
+ print(t["text"])
81
+
82
  # --- classify and route ---
83
  image_type, logo_detection = classify_image_type(detections, img_w, img_h)
84
 
 
124
  return layers, groups
125
 
126
  # --- not-logo path ---
127
+ clean_objects = run_proposal_engine(
128
+ object_proposals,
129
+ text_blocks,
130
+ img_w,
131
+ img_h,
132
+ )
133
  print(f"[LAYERS] {len(clean_objects)} objects passed proposal engine")
134
 
135
  layers = []
 
203
  ]
204
 
205
  clean_text, rejected_text = reject_text_inside_objects(text_proposals, object_boxes)
206
+
207
  if rejected_text:
208
  print(f"[LAYERS] Rejected {len(rejected_text)} text blocks inside object layers")
209
 
main.py CHANGED
@@ -6,6 +6,7 @@ from fastapi import FastAPI, File, UploadFile, HTTPException, Depends
6
  from fastapi.middleware.cors import CORSMiddleware
7
  from fastapi.responses import JSONResponse
8
  from sqlalchemy.orm import Session
 
9
 
10
  from config import UPLOAD_DIR, PROJECT_NAME
11
  from database import Base, engine, get_db
@@ -18,7 +19,8 @@ from crud import (
18
  get_all_projects,
19
  get_project_by_file_id,
20
  save_project_state,
21
- get_saved_state
 
22
  )
23
  from pipeline import run_pipeline
24
 
@@ -96,13 +98,17 @@ def list_projects(db: Session = Depends(get_db)):
96
  projects = get_all_projects(db)
97
  return [
98
  {
99
- "id": p.id,
100
- "filename": p.filename,
101
- "file_id": p.file_id,
102
- "layer_count": p.layer_count,
103
- "status": p.status,
104
- "processing_ms": p.processing_ms,
105
- "created_at": str(p.created_at),
 
 
 
 
106
  }
107
  for p in projects
108
  ]
@@ -121,7 +127,11 @@ def get_project(file_id: str, db: Session = Depends(get_db)):
121
  "layer_count": project.layer_count,
122
  "status": project.status,
123
  "processing_ms": project.processing_ms,
124
- "created_at": str(project.created_at),
 
 
 
 
125
  }
126
 
127
 
@@ -229,4 +239,32 @@ async def load_saved_project(file_id: str, db: Session = Depends(get_db)):
229
  data = get_saved_state(db, file_id)
230
  if not data:
231
  raise HTTPException(404, "Project not found")
232
- return JSONResponse(data)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6
  from fastapi.middleware.cors import CORSMiddleware
7
  from fastapi.responses import JSONResponse
8
  from sqlalchemy.orm import Session
9
+ from datetime import timezone, timedelta
10
 
11
  from config import UPLOAD_DIR, PROJECT_NAME
12
  from database import Base, engine, get_db
 
19
  get_all_projects,
20
  get_project_by_file_id,
21
  save_project_state,
22
+ get_saved_state,
23
+ delete_project,
24
  )
25
  from pipeline import run_pipeline
26
 
 
98
  projects = get_all_projects(db)
99
  return [
100
  {
101
+ "id": p.id,
102
+ "filename": p.filename,
103
+ "file_id": p.file_id,
104
+ "layer_count": p.layer_count,
105
+ "status": p.status,
106
+ "processing_ms": p.processing_ms,
107
+ "created_at": (p.created_at.replace(tzinfo=timezone.utc)
108
+ .astimezone(timezone(timedelta(hours=5, minutes=30)))
109
+ .isoformat()
110
+ ),
111
+ "thumbnail_b64": p.thumbnail_b64,
112
  }
113
  for p in projects
114
  ]
 
127
  "layer_count": project.layer_count,
128
  "status": project.status,
129
  "processing_ms": project.processing_ms,
130
+ "created_at": (
131
+ project.created_at.replace(tzinfo=timezone.utc)
132
+ .astimezone(timezone(timedelta(hours=5, minutes=30)))
133
+ .isoformat()
134
+ )
135
  }
136
 
137
 
 
239
  data = get_saved_state(db, file_id)
240
  if not data:
241
  raise HTTPException(404, "Project not found")
242
+ return JSONResponse(data)
243
+
244
+ @app.delete("/projects/{file_id}")
245
+ async def remove_project(file_id: str, db: Session = Depends(get_db)):
246
+
247
+ project = get_project_by_file_id(db, file_id)
248
+
249
+ if not project:
250
+ raise HTTPException(404, "Project not found")
251
+
252
+ # Delete uploaded image
253
+ if project.upload_path and os.path.exists(project.upload_path):
254
+ try:
255
+ os.remove(project.upload_path)
256
+ except:
257
+ pass
258
+
259
+ # Delete generated inpaint image (if present)
260
+ inpaint_path = os.path.join("outputs", f"inpaint_{file_id}.jpg")
261
+ if os.path.exists(inpaint_path):
262
+ try:
263
+ os.remove(inpaint_path)
264
+ except:
265
+ pass
266
+
267
+ # Delete database record LAST
268
+ delete_project(db, file_id)
269
+
270
+ return {"status": "deleted"}
ocr.py CHANGED
@@ -148,7 +148,7 @@ if __name__ == "__main__":
148
  TEST_IMAGE = os.path.join(
149
  BACKEND_DIR,
150
  "test_images",
151
- "sale_img.jpg"
152
  )
153
 
154
  if not os.path.exists(TEST_IMAGE):
 
148
  TEST_IMAGE = os.path.join(
149
  BACKEND_DIR,
150
  "test_images",
151
+ "logo_company.jpg"
152
  )
153
 
154
  if not os.path.exists(TEST_IMAGE):
proposal_engine.py CHANGED
@@ -91,13 +91,26 @@ def container_filter(proposals):
91
 
92
 
93
  def reject_text_inside_objects(text_proposals, accepted_objects):
94
- kept, rejected = [], []
95
- for t in text_proposals:
96
- inside = any(is_fully_inside(t, obj) for obj in accepted_objects)
97
- if inside:
98
- rejected.append((t, "inside object"))
 
 
 
 
 
 
 
 
 
 
 
 
99
  else:
100
- kept.append(t)
 
101
  return kept, rejected
102
 
103
 
@@ -150,14 +163,44 @@ def print_proposal_report(accepted, all_rejected):
150
  print(f"[PROPOSAL] {label:<25} Rejected ({reason})")
151
  print("[PROPOSAL] ─────────────────────────────────────\n")
152
 
 
 
 
 
 
 
 
 
 
 
 
 
 
153
 
154
- def run_proposal_engine(object_proposals, img_w, img_h):
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
155
  all_rejected = []
156
 
157
  after_semantic, rej = semantic_filter(object_proposals)
 
158
  all_rejected.extend(rej)
159
 
160
- after_area, rej = area_filter(after_semantic, img_w, img_h)
161
  all_rejected.extend(rej)
162
 
163
  after_dedup, rej = deduplicate(after_area)
@@ -168,4 +211,15 @@ def run_proposal_engine(object_proposals, img_w, img_h):
168
 
169
  print_proposal_report(after_container, all_rejected)
170
 
171
- return after_container
 
 
 
 
 
 
 
 
 
 
 
 
91
 
92
 
93
  def reject_text_inside_objects(text_proposals, accepted_objects):
94
+ kept = []
95
+ rejected = []
96
+
97
+ for text in text_proposals:
98
+ text_area = text["w"] * text["h"]
99
+
100
+ remove = False
101
+
102
+ for obj in accepted_objects:
103
+ overlap = intersection_area(text, obj)
104
+
105
+ if overlap / text_area >= 0.90:
106
+ remove = True
107
+ break
108
+
109
+ if remove:
110
+ rejected.append((text, "inside object"))
111
  else:
112
+ kept.append(text)
113
+
114
  return kept, rejected
115
 
116
 
 
163
  print(f"[PROPOSAL] {label:<25} Rejected ({reason})")
164
  print("[PROPOSAL] ─────────────────────────────────────\n")
165
 
166
+ def reject_text_like_objects(object_proposals, text_proposals):
167
+ kept = []
168
+ rejected = []
169
+
170
+ for obj in object_proposals:
171
+ obj_area = obj["w"] * obj["h"]
172
+ remove = False
173
+
174
+ for text in text_proposals:
175
+ overlap = intersection_area(obj, text)
176
+
177
+ if overlap == 0:
178
+ continue
179
 
180
+ text_area = text["w"] * text["h"]
181
+
182
+ text_inside = overlap / text_area
183
+ object_is_text = overlap / obj_area
184
+
185
+ if text_inside > 0.90 and object_is_text > 0.70:
186
+ remove = True
187
+ break
188
+
189
+ if remove:
190
+ rejected.append((obj, "text object"))
191
+ else:
192
+ kept.append(obj)
193
+
194
+ return kept, rejected
195
+
196
+ def run_proposal_engine(object_proposals, text_proposals, img_w, img_h):
197
  all_rejected = []
198
 
199
  after_semantic, rej = semantic_filter(object_proposals)
200
+ after_text, rej = reject_text_like_objects(after_semantic, text_proposals)
201
  all_rejected.extend(rej)
202
 
203
+ after_area, rej = area_filter(after_text, img_w, img_h)
204
  all_rejected.extend(rej)
205
 
206
  after_dedup, rej = deduplicate(after_area)
 
211
 
212
  print_proposal_report(after_container, all_rejected)
213
 
214
+ return after_container
215
+
216
+ def intersection_area(a, b):
217
+ x1 = max(a["x"], b["x"])
218
+ y1 = max(a["y"], b["y"])
219
+ x2 = min(a["x"] + a["w"], b["x"] + b["w"])
220
+ y2 = min(a["y"] + a["h"], b["y"] + b["h"])
221
+
222
+ if x2 <= x1 or y2 <= y1:
223
+ return 0
224
+
225
+ return (x2 - x1) * (y2 - y1)