utils.py

import cv2
import numpy as np
from PIL import Image
import base64
from io import BytesIO

PEAK_SMOOTHING_KERNEL, PEAK_MIN_HEIGHT_PCT, PEAK_MIN_DISTANCE = 11, 0.005, 15
PEAK_EDGE_REGION, PEAK_LOCAL_WINDOW, PEAK_EDGE_MIN_MASS_PCT = 50, 5, 0.002
CANNY_LOW, CANNY_HIGH, CANNY_DARK_THRESH, CANNY_BRIGHT_THRESH = 50, 150, 80, 180
CANNY_EDGE_RATIO_THRESHOLD, REGION_DARK_THRESHOLD, REGION_BRIGHT_THRESHOLD = 0.15, 60, 200
XDOG_SIGMA, XDOG_K, XDOG_PHI, XDOG_TAU, XDOG_EPSILON = 0.5, 1.6, 200, 0.98, 0.01
TEMPLATE_SCALE_RANGE, TEMPLATE_PYRAMID_LEVELS = (0.7, 1.3), 5
TEMPLATE_COARSE_STEPS, TEMPLATE_FINE_STEPS, TEMPLATE_TOP_CANDIDATES = 20, 10, 5

# FIX: Added min-height and box-sizing to prevent layout shifts
EMPTY = "<div style='min-height:550px;height:550px;padding:30px;color:#999;text-align:center;display:flex;align-items:center;justify-content:center;box-sizing:border-box;'>Waiting...</div>"

CONFIG = {
    'detector_type': 'orb',
    'orb_features': 5000,
    'use_template_matching': True,
    'use_gms': True,
    'use_guided_matching': True,
    'guided_search_mode': 'percent',
    'guided_search_radius_pct': 8.0,
    'guided_search_radius_px': 80,
    'bidirectional_matching': True,
    'min_guided_matches': 8,
    'gms_scale': True,
    'keypoint_mask_radius': 2,
    'ransac_iterations': 10000,
    'ransac_threshold': 5.0,
    'min_ransac_sample_distance': 30,
    'min_inliers': 4,
    'scale_min': 0.3,
    'scale_max': 3.0,
}

def get_config(key, default=None):
    return CONFIG.get(key, default)

def check_gms_available():
    try:
        return hasattr(cv2, 'xfeatures2d') and hasattr(cv2.xfeatures2d, 'matchGMS')
    except:
        return False

GMS_AVAILABLE = check_gms_available()

class Size:
    def __init__(self, width, height):
        self.width, self.height = width, height

def detect_edge_intensities(gray):
    canny = cv2.Canny(gray, CANNY_LOW, CANNY_HIGH)
    edge_count = np.sum(canny > 0)
    if edge_count < 50:
        return {'has_edges': False, 'dark_edges': False, 'bright_edges': False}, canny
    edge_y, edge_x = np.where(canny > 0)
    edge_intensities = gray[edge_y, edge_x]
    dark_count = np.sum(edge_intensities < CANNY_DARK_THRESH)
    bright_count = np.sum(edge_intensities > CANNY_BRIGHT_THRESH)
    dark_ratio, bright_ratio = dark_count / edge_count, bright_count / edge_count
    return {
        'has_edges': True,
        'dark_edges': dark_ratio > CANNY_EDGE_RATIO_THRESHOLD,
        'bright_edges': bright_ratio > CANNY_EDGE_RATIO_THRESHOLD,
        'dark_intensity': float(np.median(edge_intensities[edge_intensities < CANNY_DARK_THRESH])) if dark_count > 20 else None,
        'bright_intensity': float(np.median(edge_intensities[edge_intensities > CANNY_BRIGHT_THRESH])) if bright_count > 20 else None
    }, canny

def get_background_color(gray, canny_edges):
    non_edge_mask = canny_edges == 0
    if not np.any(non_edge_mask):
        return 255
    non_edge_pixels = gray[non_edge_mask]
    return int(np.median(non_edge_pixels)) if len(non_edge_pixels) >= 100 else 255

def analyze_color_distribution(gray):
    hist = cv2.calcHist([gray], [0], None, [256], [0, 256]).flatten()
    total = hist.sum()
    if total == 0:
        return True, 1, "empty", [], 255
    edge_info, canny = detect_edge_intensities(gray)
    bg_color = get_background_color(gray, canny)
    hist_smooth = cv2.GaussianBlur(hist.reshape(1, -1).astype(np.float32), (1, PEAK_SMOOTHING_KERNEL), 0).flatten()
    peaks = []
    if edge_info.get('dark_edges', False) and edge_info.get('dark_intensity') is not None:
        dark_mass = np.sum(hist[:PEAK_EDGE_REGION])
        peaks.append((int(edge_info['dark_intensity']), dark_mass, 'dark'))
    if edge_info.get('bright_edges', False) and edge_info.get('bright_intensity') is not None:
        bright_mass = np.sum(hist[256-PEAK_EDGE_REGION:])
        peaks.append((int(edge_info['bright_intensity']), bright_mass, 'bright'))
    min_height = total * PEAK_MIN_HEIGHT_PCT
    for i in range(PEAK_EDGE_REGION, 256 - PEAK_EDGE_REGION):
        if hist_smooth[i] < min_height:
            continue
        is_peak = all(hist_smooth[i] >= hist_smooth[max(0, i-j)] and hist_smooth[i] >= hist_smooth[min(255, i+j)] for j in range(1, PEAK_LOCAL_WINDOW + 1))
        if is_peak and not any(abs(p[0] - i) < PEAK_MIN_DISTANCE for p in peaks):
            peaks.append((i, hist_smooth[i], 'middle'))
    peaks.sort(key=lambda x: x[0])
    n = len(peaks)
    if n <= 1:
        return True, max(1, n), "single_peak", peaks, bg_color
    has_dark = any(p[0] < REGION_DARK_THRESHOLD for p in peaks)
    has_bright = any(p[0] > REGION_BRIGHT_THRESHOLD for p in peaks)
    has_middle = any(REGION_DARK_THRESHOLD <= p[0] <= REGION_BRIGHT_THRESHOLD for p in peaks)
    if has_dark and has_bright and not has_middle:
        return True, 2, "bimodal", peaks, bg_color
    return False, n, f"multicolor_{n}peaks", peaks, bg_color

def xdog_edge_detection(img, sigma=None, phi=None, tau=None):
    sigma = sigma or XDOG_SIGMA
    phi = phi or XDOG_PHI
    tau = tau or XDOG_TAU
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) if len(img.shape) == 3 else img.copy()
    gray = gray.astype(np.float64) / 255.0
    ksize1 = max(3, int(np.ceil(sigma * 3)) * 2 + 1)
    ksize2 = max(3, int(np.ceil(sigma * XDOG_K * 3)) * 2 + 1)
    g1 = cv2.GaussianBlur(gray, (ksize1, ksize1), sigma)
    g2 = cv2.GaussianBlur(gray, (ksize2, ksize2), sigma * XDOG_K)
    dog = g1 - tau * g2
    if dog.max() != dog.min():
        dog = dog / (dog.max() - dog.min() + XDOG_EPSILON)
    xdog = np.where(dog >= 0, 1.0, 1.0 + np.tanh(phi * dog))
    xdog = ((1.0 - xdog) * 255).astype(np.uint8)
    _, binary = cv2.threshold(xdog, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
    return binary

def preprocess_lineart(img):
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) if len(img.shape) == 3 else img.copy()
    is_bimodal, n_colors, reason, peaks, bg_color = analyze_color_distribution(gray)
    if is_bimodal and n_colors <= 2:
        _, binary = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
        if np.mean(binary) / 255.0 > 0.5:
            binary = cv2.bitwise_not(binary)
        return binary, "otsu", bg_color
    binary = xdog_edge_detection(img)
    if np.sum(binary > 0) / binary.size > 0.5:
        binary = cv2.bitwise_not(binary)
    return binary, "xdog", bg_color

def preprocess_color(img):
    return xdog_edge_detection(img)

def hex_to_rgb(h):
    if h is None:
        return (255, 255, 255)
    if isinstance(h, (tuple, list)):
        return tuple(int(x) for x in h[:3])
    h = str(h).strip().lstrip('#')
    if len(h) == 3:
        h = ''.join([c*2 for c in h])
    if len(h) != 6:
        return (255, 255, 255)
    try:
        return tuple(int(h[i:i+2], 16) for i in (0, 2, 4))
    except:
        return (255, 255, 255)

def flatten_rgba_to_rgb(rgba, bg):
    if len(rgba.shape) == 2:
        return cv2.cvtColor(rgba, cv2.COLOR_GRAY2RGB)
    if rgba.shape[2] == 3:
        return rgba.copy()
    alpha = rgba[:, :, 3:4].astype(np.float32) / 255.0
    rgb = rgba[:, :, :3].astype(np.float32)
    bg_array = np.array(bg, dtype=np.float32).reshape(1, 1, 3)
    return (rgb * alpha + bg_array * (1.0 - alpha)).astype(np.uint8)

def cv2_to_pil(img):
    return Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB)) if img is not None else None

def img_to_b64(img):
    buf = BytesIO()
    Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB)).save(buf, format="PNG")
    return base64.b64encode(buf.getvalue()).decode()

def make_side_by_side(a, b, max_h=500, interp=cv2.INTER_LANCZOS4):
    target_h = min(max_h, max(a.shape[0], b.shape[0]))
    s1, s2 = target_h / a.shape[0], target_h / b.shape[0]
    r1 = cv2.resize(a, (int(a.shape[1]*s1), int(a.shape[0]*s1)), interpolation=interp)
    r2 = cv2.resize(b, (int(b.shape[1]*s2), int(b.shape[0]*s2)), interpolation=interp)
    h = max(r1.shape[0], r2.shape[0])
    if r1.shape[0] < h:
        r1 = np.vstack([r1, np.ones((h-r1.shape[0], r1.shape[1], 3), np.uint8)*255])
    if r2.shape[0] < h:
        r2 = np.vstack([r2, np.ones((h-r2.shape[0], r2.shape[1], 3), np.uint8)*255])
    return np.hstack([r1, np.ones((h, 10, 3), np.uint8)*200, r2])

def make_svg(a, b, kp1, kp2, matches, max_m=200, interp=cv2.INTER_LANCZOS4):
    # FIX: Define consistent container height
    container_height = 550
    
    if not matches:
        # FIX: Return container with same fixed height as content containers
        return f"<div style='min-height:{container_height}px;height:{container_height}px;padding:30px;color:#999;text-align:center;display:flex;align-items:center;justify-content:center;box-sizing:border-box;'>No matches</div>"
    
    m, total = matches[:max_m], len(matches)
    target_h = min(500, max(a.shape[0], b.shape[0]))
    s1, s2 = target_h / a.shape[0], target_h / b.shape[0]
    h1, w1 = int(a.shape[0] * s1), int(a.shape[1] * s1)
    h2, w2 = int(b.shape[0] * s2), int(b.shape[1] * s2)
    gap, final_h, total_w = 10, max(h1, h2), w1 + 10 + w2
    a_r = cv2.resize(a, (w1, h1), interpolation=interp)
    b_r = cv2.resize(b, (w2, h2), interpolation=interp)
    np.random.seed(42)
    c = np.random.randint(50, 255, (len(m), 3))
    lines = ''.join(f'<line x1="{kp1[x.queryIdx].pt[0]*s1:.1f}" y1="{kp1[x.queryIdx].pt[1]*s1:.1f}" x2="{kp2[x.trainIdx].pt[0]*s2 + w1 + gap:.1f}" y2="{kp2[x.trainIdx].pt[1]*s2:.1f}" stroke="rgb({c[i,0]},{c[i,1]},{c[i,2]})" stroke-width="1" opacity="0.7"/>' for i, x in enumerate(m))
    circles1 = ''.join(f'<circle cx="{kp1[x.queryIdx].pt[0]*s1:.1f}" cy="{kp1[x.queryIdx].pt[1]*s1:.1f}" r="3" fill="rgb({c[i,0]},{c[i,1]},{c[i,2]})"/>' for i, x in enumerate(m))
    circles2 = ''.join(f'<circle cx="{kp2[x.trainIdx].pt[0]*s2 + w1 + gap:.1f}" cy="{kp2[x.trainIdx].pt[1]*s2:.1f}" r="3" fill="rgb({c[i,0]},{c[i,1]},{c[i,2]})"/>' for i, x in enumerate(m))
    info = f"{len(m)} of {total} matches" if len(m) < total else f"{total} matches"
    
    # FIX: Wrap in container with fixed height and use preserveAspectRatio
    return f'''<div style="min-height:{container_height}px;height:{container_height}px;box-sizing:border-box;overflow:hidden;">
<div style="font-size:12px;color:#666;margin-bottom:4px;height:20px;">{info}</div>
<svg viewBox="0 0 {total_w} {final_h}" preserveAspectRatio="xMidYMid meet" style="width:100%;height:{min(final_h, container_height - 30)}px;max-height:{container_height - 30}px;background:#fafafa;display:block;">
<image href="data:image/png;base64,{img_to_b64(a_r)}" width="{w1}" height="{h1}"/>
<rect x="{w1}" y="0" width="{gap}" height="{final_h}" fill="#e0e0e0"/>
<image x="{w1 + gap}" href="data:image/png;base64,{img_to_b64(b_r)}" width="{w2}" height="{h2}"/>
{lines}{circles1}{circles2}
</svg></div>'''

def edge_filter(matches, kp2, edges, r):
    h, w, out = edges.shape[0], edges.shape[1], []
    for m in matches:
        y, x = int(kp2[m.trainIdx].pt[1]), int(kp2[m.trainIdx].pt[0])
        if np.any(edges[max(0,y-r):min(h,y+r+1), max(0,x-r):min(w,x+r+1)]):
            out.append(m)
    return out

def filter_keypoints_by_mask(keypoints, descriptors, mask, radius=2):
    if keypoints is None or descriptors is None or len(keypoints) == 0:
        return [], None
    h, w = mask.shape[:2]
    keep_idx = []
    for i, kp in enumerate(keypoints):
        x, y = int(round(kp.pt[0])), int(round(kp.pt[1]))
        x1, y1 = max(0, x - radius), max(0, y - radius)
        x2, y2 = min(w, x + radius + 1), min(h, y + radius + 1)
        if x2 > x1 and y2 > y1 and np.any(mask[y1:y2, x1:x2] > 0):
            keep_idx.append(i)
    if not keep_idx:
        return list(keypoints), descriptors
    return [keypoints[i] for i in keep_idx], descriptors[keep_idx]

def visualize_template_match(lineart, color, scale, tx, ty, score):
    h, w = color.shape[:2]
    nw, nh = int(lineart.shape[1] * scale), int(lineart.shape[0] * scale)
    if nw <= 0 or nh <= 0:
        return color.copy()
    scaled = cv2.resize(lineart, (nw, nh), interpolation=cv2.INTER_LANCZOS4)
    overlay = color.copy()
    txi, tyi = int(tx), int(ty)
    dx1, dy1 = max(0, txi), max(0, tyi)
    dx2, dy2 = min(w, txi + nw), min(h, tyi + nh)
    sx1, sy1 = max(0, -txi), max(0, -tyi)
    if dy2 > dy1 and dx2 > dx1:
        h_copy, w_copy = dy2 - dy1, dx2 - dx1
        roi = overlay[dy1:dy2, dx1:dx2]
        scaled_roi = scaled[sy1:sy1+h_copy, sx1:sx1+w_copy]
        if len(scaled_roi.shape) == 2:
            scaled_roi = cv2.cvtColor(scaled_roi, cv2.COLOR_GRAY2BGR)
        overlay[dy1:dy2, dx1:dx2] = cv2.addWeighted(roi, 0.5, scaled_roi, 0.5, 0)
        cv2.rectangle(overlay, (dx1, dy1), (dx2-1, dy2-1), (0, 255, 0), 2)
    font = cv2.FONT_HERSHEY_SIMPLEX
    text = f"Template: scale={scale:.3f} tx={tx:.0f} ty={ty:.0f} score={score:.3f}"
    cv2.putText(overlay, text, (10, 30), font, 0.7, (0, 255, 0), 2)
    return overlay

def resize_image_match(finished_img, lineart_size):
    finished_width, finished_height = finished_img.size
    lineart_width, lineart_height = lineart_size
    finished_longest = max(finished_width, finished_height)
    finished_shortest = min(finished_width, finished_height)
    lineart_longest = max(lineart_width, lineart_height)
    lineart_shortest = min(lineart_width, lineart_height)
    if finished_longest > lineart_longest:
        scale = lineart_longest / finished_longest
    else:
        scale = lineart_shortest / finished_shortest
    new_width = int(finished_width * scale)
    new_height = int(finished_height * scale)
    return finished_img.resize((new_width, new_height), Image.LANCZOS)

def get_guided_search_radius(w, h, config):
    if config.get('guided_search_mode', 'percent') == 'percent':
        return max(w, h) * config.get('guided_search_radius_pct', 8.0) / 100.0
    return config.get('guided_search_radius_px', 80)