File size: 21,678 Bytes
4752ca3
 
de2a7cd
 
 
 
4752ca3
 
de2a7cd
4752ca3
 
 
 
 
8501481
 
de2a7cd
b57f874
de2a7cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b57f874
de2a7cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6ad3a74
fcccc26
6ad3a74
b57f874
9a51cf4
b57f874
 
 
 
 
 
fcccc26
7412522
fcccc26
 
 
 
7412522
fcccc26
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7412522
 
de2a7cd
b57f874
 
9a51cf4
 
 
de2a7cd
 
4752ca3
fcccc26
4752ca3
 
b57f874
 
 
de2a7cd
4752ca3
de2a7cd
4752ca3
 
9a51cf4
b57f874
 
 
 
 
4752ca3
 
 
 
 
fcccc26
7412522
6ad3a74
fcccc26
 
 
4752ca3
9a51cf4
de2a7cd
 
 
 
 
 
4752ca3
fcccc26
 
4752ca3
de2a7cd
fcccc26
 
 
 
 
de2a7cd
 
b57f874
de2a7cd
 
fcccc26
de2a7cd
fcccc26
7412522
fcccc26
7412522
 
 
6ad3a74
9a51cf4
de2a7cd
b57f874
9a51cf4
 
b57f874
de2a7cd
 
 
 
 
 
 
9a51cf4
de2a7cd
 
 
 
 
 
b57f874
de2a7cd
 
 
 
 
 
 
 
b57f874
 
 
 
de2a7cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b57f874
de2a7cd
 
 
 
 
 
 
b57f874
 
de2a7cd
 
9a51cf4
 
 
 
b57f874
de2a7cd
 
 
b57f874
 
de2a7cd
 
 
b57f874
 
 
 
de2a7cd
 
 
 
 
 
 
b57f874
 
 
 
 
 
 
 
 
 
 
de2a7cd
 
 
 
 
 
b57f874
de2a7cd
 
 
 
 
 
b57f874
 
 
 
 
 
 
 
de2a7cd
 
 
 
 
 
 
 
 
b57f874
 
de2a7cd
 
 
 
 
 
 
 
b57f874
 
 
 
 
 
 
 
 
 
 
de2a7cd
 
4752ca3
de2a7cd
 
9a51cf4
 
 
de2a7cd
 
 
 
 
 
 
 
9a51cf4
de2a7cd
9a51cf4
de2a7cd
 
9a51cf4
 
 
 
 
 
 
 
 
 
 
 
 
de2a7cd
 
9a51cf4
4752ca3
 
de2a7cd
 
 
 
 
 
 
 
 
4752ca3
 
 
de2a7cd
4752ca3
de2a7cd
b57f874
 
9a51cf4
 
de2a7cd
 
 
4752ca3
de2a7cd
8501481
 
de2a7cd
8501481
4752ca3
 
de2a7cd
 
4752ca3
de2a7cd
4752ca3
 
de2a7cd
 
 
 
4752ca3
 
de2a7cd
 
9a51cf4
 
de2a7cd
4752ca3
 
de2a7cd
 
 
 
 
b57f874
9a51cf4
 
b57f874
de2a7cd
 
b57f874
9a51cf4
 
b57f874
de2a7cd
 
 
4752ca3
b57f874
de2a7cd
b57f874
4752ca3
 
 
 
9a51cf4
 
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
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
import gradio as gr
import cv2
import numpy as np
import pandas as pd
import concurrent.futures
from PIL import Image, ImageDraw, ImageFont
from ultralytics import YOLO

# ─── Model ────────────────────────────────────────────────────────────────────
model = YOLO("best.pt")

CLASS_NAMES  = {0: "Full",   1: "Broken"}
CLASS_COLORS = {0: (34, 197, 94), 1: (239, 68, 68)}   # green, red

SAMPLE_PATHS = ["image1.jpg", "image2.jpg"]

# ─── Paper reference ──────────────────────────────────────────────────────────
PAPER_REAL_MM = 40.0   # white 4x4 cm square = 40 mm per side

def detect_paper_pixels(img_np):
    gray = cv2.cvtColor(img_np, cv2.COLOR_RGB2GRAY)
    _, thresh = cv2.threshold(gray, 180, 255, cv2.THRESH_BINARY)
    kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (7, 7))
    thresh = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, kernel)
    thresh = cv2.morphologyEx(thresh, cv2.MORPH_OPEN,  kernel)
    contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    img_area  = img_np.shape[0] * img_np.shape[1]
    best, best_area = None, 0
    for c in contours:
        area = cv2.contourArea(c)
        if area < img_area * 0.02:
            continue
        x, y, w, h = cv2.boundingRect(c)
        if 0.5 < (w / max(h, 1)) < 2.0 and area > best_area:
            best_area = area
            best = (h, w)
    return best

def px_to_mm(px, paper_px_dim):
    if not paper_px_dim:
        return None
    return px * PAPER_REAL_MM / paper_px_dim

# ─── Font helper ──────────────────────────────────────────────────────────────
def _font(size, bold=False):
    for path in [
        "/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf" if bold else
        "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf",
        "/usr/share/fonts/truetype/liberation/LiberationSans-Regular.ttf",
    ]:
        try:
            return ImageFont.truetype(path, size)
        except Exception:
            pass
    return ImageFont.load_default()

def _text_size(draw, text, font):
    bbox = draw.textbbox((0, 0), text, font=font)
    return bbox[2] - bbox[0], bbox[3] - bbox[1]


# ─── Mask helpers ─────────────────────────────────────────────────────────────

def _polygon_to_mask(pts_xy, h, w):
    """Rasterise raw polygon β†’ binary uint8 mask. BACKEND / measurements only."""
    mask = np.zeros((h, w), dtype=np.uint8)
    if len(pts_xy) >= 3:
        cv2.fillPoly(mask, [pts_xy.astype(np.int32)], 1)
    return mask


def _refine_mask_grabcut(img_bgr, coarse_mask):
    """
    Refine a coarse binary mask to pixel-perfect grain boundary using GrabCut.
    img_bgr     : full BGR image
    coarse_mask : uint8 binary mask (0/1), same size as img_bgr
    Returns     : refined binary uint8 mask (0/1)
    """
    ys, xs = np.where(coarse_mask == 1)
    if len(xs) < 5:
        return coarse_mask

    # Tight crop with small padding so GrabCut has background context
    x1, y1 = max(0, int(xs.min()) - 6), max(0, int(ys.min()) - 6)
    x2, y2 = min(img_bgr.shape[1], int(xs.max()) + 6), min(img_bgr.shape[0], int(ys.max()) + 6)
    crop    = img_bgr[y1:y2, x1:x2]
    ch, cw  = crop.shape[:2]
    if ch < 8 or cw < 8:
        return coarse_mask

    # Build GrabCut init mask from coarse mask crop
    gc_mask  = np.full((ch, cw), cv2.GC_BGD, dtype=np.uint8)
    local_fg = coarse_mask[y1:y2, x1:x2]

    # Erode to get definite FG core, dilate to get probable FG ring
    k_sm = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
    k_lg = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (11, 11))
    def_fg   = cv2.erode(local_fg,  k_sm, iterations=2)
    prob_fg  = cv2.dilate(local_fg, k_lg, iterations=2)

    gc_mask[prob_fg == 1] = cv2.GC_PR_FGD
    gc_mask[def_fg  == 1] = cv2.GC_FGD
    # Border strip = definite background
    gc_mask[:3, :]  = cv2.GC_BGD
    gc_mask[-3:, :] = cv2.GC_BGD
    gc_mask[:, :3]  = cv2.GC_BGD
    gc_mask[:, -3:] = cv2.GC_BGD

    try:
        bgd_model = np.zeros((1, 65), np.float64)
        fgd_model = np.zeros((1, 65), np.float64)
        cv2.grabCut(crop, gc_mask, None, bgd_model, fgd_model, 4, cv2.GC_INIT_WITH_MASK)
        refined_local = np.where((gc_mask == cv2.GC_FGD) | (gc_mask == cv2.GC_PR_FGD), 1, 0).astype(np.uint8)
    except Exception:
        return coarse_mask

    # Clean up with morphology: close small holes, smooth jagged edges
    k_cl = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
    refined_local = cv2.morphologyEx(refined_local, cv2.MORPH_CLOSE, k_cl, iterations=2)
    refined_local = cv2.morphologyEx(refined_local, cv2.MORPH_OPEN,  k_cl, iterations=1)

    # Put refined crop back into full-size mask
    refined_full = np.zeros_like(coarse_mask)
    refined_full[y1:y2, x1:x2] = refined_local
    return refined_full


def _mask_to_smooth_contour(mask_np):
    """
    Extract the outer contour of a binary mask and smooth it with
    spline-like resampling β†’ returns int32 array (N,1,2) for cv2 drawing.
    """
    contours, _ = cv2.findContours(mask_np, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
    if not contours:
        return None
    cnt = max(contours, key=cv2.contourArea).astype(np.float32).reshape(-1, 2)
    if len(cnt) < 6:
        return cnt.astype(np.int32).reshape(-1, 1, 2)

    # Resample to ~120 evenly-spaced points for a smooth outline
    n_target = min(120, max(40, len(cnt)))
    indices  = np.linspace(0, len(cnt) - 1, n_target).astype(int)
    sampled  = cnt[indices]

    # Circular Gaussian smooth
    window   = 9
    half     = window // 2
    padded   = np.vstack([sampled[-half:], sampled, sampled[:half]])
    kernel   = cv2.getGaussianKernel(window, 0).flatten()
    kernel  /= kernel.sum()
    smoothed = np.zeros_like(sampled)
    for i in range(len(sampled)):
        smoothed[i] = (padded[i:i + window] * kernel[:, None]).sum(axis=0)

    return smoothed.astype(np.int32).reshape(-1, 1, 2)


# ─────────────────────────────────────────────────────────────────────────────
# STEP 1 β€” Segmentation + visual output
#
# Uses results.masks.xy (polygon in original-image px coords) instead of
# results.masks.data (low-res tensor + resize) β†’ zero resize drift,
# pixel-perfect mask alignment.
# ─────────────────────────────────────────────────────────────────────────────
def run_segmentation(img_np):
    h, w    = img_np.shape[:2]
    img_bgr = cv2.cvtColor(img_np, cv2.COLOR_RGB2BGR)
    results = model(img_np, imgsz=1280, conf=0.25)[0]

    annotated   = img_np.copy()
    overlay     = img_np.copy()
    counts      = {"Full": 0, "Broken": 0}
    grain_boxes = []

    all_x1, all_y1, all_x2, all_y2 = w, h, 0, 0

    if results.masks is not None:
        xy_list = results.masks.xy      # list of (N_i, 2) float arrays, orig coords

        for poly_xy, box in zip(xy_list, results.boxes):
            if len(poly_xy) < 3:
                continue

            cls_id   = int(box.cls[0])
            cls_name = CLASS_NAMES.get(cls_id, "?")
            color    = CLASS_COLORS.get(cls_id, (200, 200, 200))
            counts[cls_name] += 1

            # Backend mask: raw polygon fill (used for measurements β€” never changed)
            mask_np = _polygon_to_mask(poly_xy, h, w)

            # Visual mask: GrabCut-refined β†’ hugs actual grain pixels perfectly
            vis_mask    = _refine_mask_grabcut(img_bgr, mask_np)
            vis_contour = _mask_to_smooth_contour(vis_mask)

            # Bounding box from backend mask for zoom crop
            ys, xs = np.where(mask_np == 1)
            if len(xs) > 0:
                all_x1 = min(all_x1, int(xs.min()))
                all_y1 = min(all_y1, int(ys.min()))
                all_x2 = max(all_x2, int(xs.max()))
                all_y2 = max(all_y2, int(ys.max()))

            # Fill overlay using the refined visual mask directly (pixel-perfect fill)
            overlay[vis_mask == 1] = color

            grain_boxes.append({
                "cls_id":     cls_id,
                "cls_name":   cls_name,
                "mask_np":    mask_np,      # backend only β€” measurements
                "vis_mask":   vis_mask,     # refined visual mask
                "vis_contour": vis_contour, # smooth contour for outline
            })

    # Blend fill
    annotated = cv2.addWeighted(annotated, 0.72, overlay, 0.28, 0)

    # Draw smooth anti-aliased contour outlines over the blend
    for g in grain_boxes:
        if g["vis_contour"] is not None:
            cv2.polylines(
                annotated, [g["vis_contour"]],
                isClosed=True, color=CLASS_COLORS[g["cls_id"]], thickness=2,
                lineType=cv2.LINE_AA,
            )

    # Zoom
    if all_x2 > all_x1 and all_y2 > all_y1:
        pad  = max(30, int(max(all_x2 - all_x1, all_y2 - all_y1) * 0.08))
        cx1, cy1 = max(0, all_x1 - pad), max(0, all_y1 - pad)
        cx2, cy2 = min(w, all_x2 + pad), min(h, all_y2 + pad)
        zoomed_pil = Image.fromarray(annotated[cy1:cy2, cx1:cx2])
    else:
        zoomed_pil = Image.fromarray(annotated)

    return annotated, zoomed_pil, grain_boxes, counts


# ─────────────────────────────────────────────────────────────────────────────
# STEP 2 β€” Measure grains (backend mask_np only β€” unaffected by visual changes)
# ─────────────────────────────────────────────────────────────────────────────
def measure_grains_from_boxes(grain_boxes, img_shape, paper_dims):
    paper_px = (paper_dims[0] + paper_dims[1]) / 2.0 if paper_dims else None
    measurements = []

    for idx, g in enumerate(grain_boxes):
        mask_np = g["mask_np"]
        pts_y, pts_x = np.where(mask_np == 1)
        if len(pts_x) < 5:
            continue

        pts  = np.column_stack([pts_x.astype(np.float32), pts_y.astype(np.float32)])
        rect = cv2.minAreaRect(pts)
        (cx, cy), (rw, rh), _ = rect

        h_px     = float(max(rw, rh))
        w_px     = float(min(rw, rh))
        h_mm     = px_to_mm(h_px, paper_px)
        w_mm     = px_to_mm(w_px, paper_px)
        area_mm2 = (h_mm * w_mm) if (h_mm and w_mm) else None

        measurements.append({
            "label":      idx + 1,
            "cls_name":   g["cls_name"],
            "h_px":       h_px,
            "w_px":       w_px,
            "h_mm":       h_mm,
            "w_mm":       w_mm,
            "area_mm2":   area_mm2,
            "centroid_x": int(cx),
            "centroid_y": int(cy),
        })

    return measurements, paper_px


# ─────────────────────────────────────────────────────────────────────────────
# STEP 2b β€” Build DataFrames
# ─────────────────────────────────────────────────────────────────────────────
def build_table_data(measurements, paper_px, counts):
    has_mm = paper_px is not None
    unit   = "mm" if has_mm else "px"

    rows = []
    for g in measurements:
        h_val    = round(g["h_mm"],    2) if (has_mm and g["h_mm"])    else round(g["h_px"], 1)
        w_val    = round(g["w_mm"],    2) if (has_mm and g["w_mm"])    else round(g["w_px"], 1)
        area_val = round(g["area_mm2"], 2) if g["area_mm2"] else None
        rows.append({
            "#":                                     g["label"],
            "Class":                                 g["cls_name"],
            f"Height ({unit})":                      h_val,
            f"Width ({unit})":                       w_val,
            "Area (mm\u00b2)" if has_mm else "Area": area_val,
        })
    grain_df = pd.DataFrame(rows)

    h_key   = "h_mm" if has_mm else "h_px"
    w_key   = "w_mm" if has_mm else "w_px"
    heights = [(g["label"], g[h_key]) for g in measurements if g.get(h_key)]
    widths  = [(g["label"], g[w_key]) for g in measurements if g.get(w_key)]

    max_h    = max(heights, key=lambda x: x[1]) if heights else (0, 0)
    min_h    = min(heights, key=lambda x: x[1]) if heights else (0, 0)
    max_w    = max(widths,  key=lambda x: x[1]) if widths  else (0, 0)
    min_w    = min(widths,  key=lambda x: x[1]) if widths  else (0, 0)
    interval = (max_h[1] - min_h[1]) / 10.0 if (heights and max_h[1] != min_h[1]) else 0.0

    n_full   = counts.get("Full",   0)
    n_broken = counts.get("Broken", 0)
    total    = n_full + n_broken

    summary_rows = [
        {"Metric": "Total Grains",                    "Value": str(total)},
        {"Metric": "🟒 Full Grains",                  "Value": str(n_full)},
        {"Metric": "πŸ”΄ Broken Grains",                "Value": str(n_broken)},
        {"Metric": "Paper Reference",                 "Value": f"βœ… Found ({unit} mode)" if has_mm else "❌ Not found (px only)"},
        {"Metric": f"Max Height (Grain #{max_h[0]})", "Value": f"{max_h[1]:.2f} {unit}"},
        {"Metric": f"Min Height (Grain #{min_h[0]})", "Value": f"{min_h[1]:.2f} {unit}"},
        {"Metric": f"Max Width  (Grain #{max_w[0]})", "Value": f"{max_w[1]:.2f} {unit}"},
        {"Metric": f"Min Width  (Grain #{min_w[0]})", "Value": f"{min_w[1]:.2f} {unit}"},
        {"Metric": "Mean Height",                     "Value": f"{np.mean([v for _, v in heights]):.2f} {unit}" if heights else "β€”"},
        {"Metric": "Mean Width",                      "Value": f"{np.mean([v for _, v in widths]):.2f} {unit}"  if widths  else "β€”"},
        {"Metric": "Bin Interval (max-min)/10",       "Value": f"{interval:.3f} {unit}"},
    ]
    summary_df = pd.DataFrame(summary_rows)
    return grain_df, summary_df


# ─────────────────────────────────────────────────────────────────────────────
# GRADIO β€” two-stage predict
# ─────────────────────────────────────────────────────────────────────────────
def predict_stage1(image: Image.Image):
    if image is None:
        return None, "", "", None, None
    img_np = np.array(image.convert("RGB"))
    _, zoomed_pil, grain_boxes, counts = run_segmentation(img_np)
    total    = counts["Full"] + counts["Broken"]
    summary  = f"βœ…  {total} grains detected  Β·  🟒 Full: {counts['Full']}  Β·  πŸ”΄ Broken: {counts['Broken']}"
    count_md = (
        f"| | Count |\n|---|---|\n"
        f"| 🌾 Total Grains | **{total}** |\n"
        f"| 🟒 Full Grains | **{counts['Full']}** |\n"
        f"| πŸ”΄ Broken Grains | **{counts['Broken']}** |"
    )
    loading_df = pd.DataFrame([{"Status": "⏳  Calculating height & width of all grains..."}])
    return zoomed_pil, summary, count_md, loading_df, loading_df


def predict_stage2(image: Image.Image):
    if image is None:
        return None, "", "", None, None
    img_np = np.array(image.convert("RGB"))

    def _seg():   return run_segmentation(img_np)
    def _paper(): return detect_paper_pixels(img_np)

    with concurrent.futures.ThreadPoolExecutor(max_workers=2) as pool:
        fut_seg   = pool.submit(_seg)
        fut_paper = pool.submit(_paper)
        _, zoomed_pil, grain_boxes, counts = fut_seg.result()
        paper_dims = fut_paper.result()

    measurements, paper_px = measure_grains_from_boxes(grain_boxes, img_np.shape, paper_dims)
    total    = counts["Full"] + counts["Broken"]
    summary  = (
        f"βœ…  {total} grains detected  Β·  🟒 Full: {counts['Full']}  Β·  πŸ”΄ Broken: {counts['Broken']}"
        + (f"  Β·  πŸ“ Paper found β€” measurements in mm" if paper_px else "  Β·  ⚠️ No paper β€” measurements in px")
    )
    count_md = (
        f"| | Count |\n|---|---|\n"
        f"| 🌾 Total Grains | **{total}** |\n"
        f"| 🟒 Full Grains | **{counts['Full']}** |\n"
        f"| πŸ”΄ Broken Grains | **{counts['Broken']}** |"
    )
    grain_df, summary_df = build_table_data(measurements, paper_px, counts)
    return zoomed_pil, summary, count_md, grain_df, summary_df


# ─────────────────────────────────────────────────────────────────────────────
# UI  β€” Gradio 6 compatible
#   β€’ theme / css β†’ moved to demo.launch()
#   β€’ gr.DataFrame has no height param β†’ use CSS to expand tables
# ─────────────────────────────────────────────────────────────────────────────
THEME = gr.themes.Soft(
    primary_hue="violet",
    secondary_hue="indigo",
    neutral_hue="slate",
    font=gr.themes.GoogleFont("Inter"),
)

# In Gradio 6 the CSS string is passed to launch(), not Blocks()
CSS = """
#run-btn  { margin-top: 6px; }
#status-box textarea { font-size: 0.92rem; }
#count-box { font-size: 0.95rem; }

/* Make both measurement tables tall enough to show all rows */
#grain-table   .table-wrap,
#grain-table   .svelte-table,
#summary-table .table-wrap,
#summary-table .svelte-table {
    max-height: none !important;
    overflow-y: visible !important;
}
#grain-table,
#summary-table {
    overflow: visible !important;
}
"""

with gr.Blocks(title="GrainVision") as demo:

    gr.HTML("""
    <div style="padding:18px 12px 10px 12px; background-color:#0F172A;
                border-radius:10px; margin-bottom:10px;">
      <span style="font-size:2rem;font-weight:900;color:#F1F5F9;font-family:sans-serif;">
        🌾 GrainVision
      </span>
      <p style="color:#CBD5E1;font-size:0.9rem;margin-top:4px;font-family:sans-serif;">
        Upload a rice image (with white 4Γ—4 cm reference paper) to segment, classify,
        measure, and analyse grains.
      </p>
    </div>
    """)

    with gr.Row(equal_height=False):
        with gr.Column(scale=1):
            inp_image = gr.Image(type="pil", label="Upload Rice Image", height=280)
            run_btn   = gr.Button("πŸ”  Analyse Grains",
                                  variant="primary", size="lg", elem_id="run-btn")
            gr.Markdown("_Upload an image then press **Analyse**. "
                        "Segmentation appears first, measurements follow._")
            status_box = gr.Textbox(
                label="Status", value="", interactive=False,
                visible=True, max_lines=3, elem_id="status-box",
            )
            gr.Markdown("### Example Images  _(click to load)_")
            gr.Examples(
                examples=[[p] for p in SAMPLE_PATHS],
                inputs=inp_image, label="", examples_per_page=6,
            )

        with gr.Column(scale=1):
            gr.Markdown("### Segmentation Output  *(zoomed to grains)*")
            seg_out = gr.Image(label="", interactive=False)

    gr.Markdown("---")
    with gr.Row():
        with gr.Column(scale=1):
            gr.Markdown("#### Detection Summary")
            summary_box = gr.Textbox(
                label="", value="", interactive=False,
                max_lines=2, elem_id="status-box",
            )
        with gr.Column(scale=1):
            gr.Markdown("#### Grain Count")
            count_md = gr.Markdown(
                value="| | Count |\n|---|---|\n"
                      "| 🌾 Total | β€” |\n| 🟒 Full | β€” |\n| πŸ”΄ Broken | β€” |",
                elem_id="count-box",
            )

    gr.Markdown("---")
    gr.Markdown("### Grain Measurements Table")
    with gr.Row():
        with gr.Column(scale=2):
            gr.Markdown("#### Per-Grain Measurements")
            grain_table_out = gr.DataFrame(
                label="", interactive=False, wrap=False,
                elem_id="grain-table",
            )
        with gr.Column(scale=1):
            gr.Markdown("#### Summary Statistics")
            summary_table_out = gr.DataFrame(
                label="", interactive=False, wrap=False,
                elem_id="summary-table",
            )

    OUTPUTS = [seg_out, summary_box, count_md, grain_table_out, summary_table_out]

    run_btn.click(
        fn=predict_stage1, inputs=[inp_image], outputs=OUTPUTS,
    ).then(
        fn=predict_stage2, inputs=[inp_image], outputs=OUTPUTS,
    )


if __name__ == "__main__":
    # Gradio 6: theme and css passed here, not in gr.Blocks()
    demo.launch(theme=THEME, css=CSS)