import sys
sys.path.insert(0, "services/recolor")
import numpy as np
from PIL import Image
import cv2

from fabric_recolor.segmentation.birefnet import BirefnetMatteSegmenter
from fabric_recolor.segmentation.grounded_sam import GroundedSamSegmenter
from fabric_recolor.segmentation.prompts import NEGATIVE_PROMPTS, CAVITY_PROMPTS
from fabric_recolor.masks import clean_binary, largest_component, fill_small_holes, alpha_hysteresis

src = "data/media/originals/fd/fd12e584a23312c7c5edda1015fe9715d3f5fd19a2ef3d375d635cefab91f54b.jpg"
img = Image.open(src).convert("RGB")
W0, H0 = img.width, img.height
scale = min(1.0, 1600 / max(H0, W0))
seg = img.resize((round(W0*scale), round(H0*scale)), Image.BILINEAR)
arr = np.asarray(seg)

seg_b = BirefnetMatteSegmenter()
seg_b._load()
matte = seg_b._matte(seg)
matte = alpha_hysteresis(matte, ramp_px=max(4, round(8*scale)))
mask = clean_binary((matte >= 0.5).astype(np.uint8), min_area_frac=0.002)
mask = largest_component(mask)
mask = fill_small_holes(mask, 0.01)
obj = int(mask.sum())

# convexity defects = wnęki sylwetki (otwarte i zamknięte)
cnts, _ = cv2.findContours(mask.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
hull_img = np.zeros_like(mask)
for c in cnts:
    cv2.fillPoly(hull_img, [cv2.convexHull(c)], 1)
defect = (hull_img & (1 - mask)).astype(np.uint8)
# podział defectów: połączone z zewnętrzem (otwarte wnęki) vs zamknięte
n_d, l_d = cv2.connectedComponents(defect, connectivity=8)
bg_reach = np.zeros_like(mask)
ff = (1 - mask).astype(np.uint8).copy()
cv2.floodFill(ff, np.zeros((ff.shape[0]+2, ff.shape[1]+2), np.uint8), (0,0), 2)
open_ids = np.unique(l_d[ff == 2]); open_ids = open_ids[open_ids > 0]
bay = np.isin(l_d, open_ids).astype(np.uint8)
print("mask:", obj, "hull:", int(hull_img.sum()), "defect:", int(defect.sum()), "bay(open):", int(bay.sum()))

# band_out: maska przyległa do tła poza hull (prawdziwy obrys)
k6 = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (13, 13))
outside = ((1 - mask) & (1 - hull_img)).astype(np.uint8)
band_out = (mask > 0) & (cv2.dilate(outside, k6) > 0)

viz = arr.copy()
viz[mask > 0] = (viz[mask > 0]*0.5 + np.array([0,180,0])*0.5).astype(np.uint8)
viz[bay > 0] = (viz[bay > 0]*0.4 + np.array([0,80,255])*0.6).astype(np.uint8)
viz[band_out] = [255, 0, 0]
Image.fromarray(viz).save("tmp/dbg-bands-6413.png")

iso = Image.fromarray(arr * mask[..., None].astype(np.uint8))
gs = GroundedSamSegmenter()
gs._load()
det = gs._detect(iso, list(NEGATIVE_PROMPTS))
labels = det.get("text_labels") or det.get("labels") or []
boxes = det["boxes"].cpu().numpy()
scores = det["scores"].cpu().numpy()
masks = gs._boxes_to_masks(iso, boxes)

for i, (lab, box, sc) in enumerate(zip(labels, boxes, scores)):
    if i >= len(masks): break
    m = masks[i]
    pm = m & (mask > 0)
    a = int(pm.sum())
    if a == 0:
        print(f"{sc:.2f} '{lab}' -> 0 inside"); continue
    bf = float((pm & band_out).sum()) / a
    bay_adj = bool((cv2.dilate(pm.astype(np.uint8), k6) & bay.astype(bool)).any())
    idx = next((j for j,p in enumerate(NEGATIVE_PROMPTS) if p in lab or lab in p), -1)
    cav = idx >= 0 and NEGATIVE_PROMPTS[idx] in CAVITY_PROMPTS
    print(f"[{i}] {sc:.2f} '{lab}' inside={a} ({a/obj:.0%}) band={bf:.2f} bay_adj={bay_adj} cav={cav}")
    v = arr.copy()
    v[pm] = (v[pm]*0.35 + np.array([255,40,40])*0.65).astype(np.uint8)
    v[m & (mask == 0)] = (v[m & (mask == 0)]*0.35 + np.array([255,200,40])*0.65).astype(np.uint8)
    Image.fromarray(v).resize((600, 800)).save(f"tmp/inst-{i}-6413.png")
print("saved tmp/inst-*-6413.png, tmp/dbg-bands-6413.png")