app.py

import gradio as gr
import torch
import numpy as np
import cv2
from PIL import Image, ImageOps
from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from paddleocr import PaddleOCR
from scipy.signal import find_peaks

# ==========================================
# ⚙️ CONFIGURATION & MODEL LOADING
# ==========================================
print("--- SYSTEM STARTUP ---")

# Force CPU to avoid CUDA overhead on CPU-only Spaces
DEVICE = "cpu"
print(f"-> Hardware Device: {DEVICE}")

# 1. LOAD TR-OCR (Recognition)
# We use the 'stage1' model which is often more robust for general handwriting
print("-> Loading TrOCR Model...")
processor = TrOCRProcessor.from_pretrained('microsoft/trocr-base-handwritten')
model = VisionEncoderDecoderModel.from_pretrained('microsoft/trocr-base-handwritten').to(DEVICE).eval()

# 2. LOAD PADDLEOCR (Detection)
# 'structure_version' and generic settings tuned for recall (catch everything, filter later)
print("-> Loading PaddleOCR Detector...")
detector = PaddleOCR(
    use_angle_cls=True, 
    lang='en', 
    show_log=False,
    use_gpu=False,
    det_limit_side_len=2500,  # High res for small text
    det_db_thresh=0.1,        # Low threshold to catch faint ink
    det_db_box_thresh=0.3,
    det_db_unclip_ratio=1.6
)
print("--- SYSTEMS READY ---")

# ==========================================
# 🧠 CORE LOGIC: GEOMETRY UTILS
# ==========================================

def calculate_iou_containment(box1, box2):
    """
    Calculates how much of box1 is inside box2.
    """
    x1 = max(box1[0], box2[0])
    y1 = max(box1[1], box2[1])
    x2 = min(box1[2], box2[2])
    y2 = min(box1[3], box2[3])
    
    if x2 < x1 or y2 < y1: return 0.0
    
    intersection = (x2 - x1) * (y2 - y1)
    area1 = (box1[2] - box1[0]) * (box1[3] - box1[1]) + 1e-6
    return intersection / area1

def get_vertical_overlap_ratio(box1, box2):
    """
    Calculates vertical overlap between two boxes.
    Used to determine if words are on the same line.
    """
    # y1, y2 are top, bottom
    y1_a, y2_a = box1[1], box1[3]
    y1_b, y2_b = box2[1], box2[3]
    
    intersection_start = max(y1_a, y1_b)
    intersection_end = min(y2_a, y2_b)
    
    if intersection_end < intersection_start: return 0.0
    
    overlap_height = intersection_end - intersection_start
    min_height = min(y2_a - y1_a, y2_b - y1_b) + 1e-6
    
    return overlap_height / min_height

def filter_nested_boxes(boxes, containment_thresh=0.9):
    """
    Removes small noise boxes inside larger real boxes.
    """
    if not boxes: return []
    
    # Add area to list: [x1, y1, x2, y2, area]
    active = []
    for b in boxes:
        area = (b[2] - b[0]) * (b[3] - b[1])
        active.append(list(b) + [area])
    
    # Sort largest first
    active.sort(key=lambda x: x[4], reverse=True)
    
    final_boxes = []
    for current in active:
        is_nested = False
        curr_box = current[:4]
        
        for kept in final_boxes:
            if calculate_iou_containment(curr_box, kept) > containment_thresh:
                is_nested = True
                break
        
        if not is_nested:
            final_boxes.append(curr_box)
            
    return final_boxes

# ==========================================
# 🔬 SCIENTIFIC LOGIC: PROJECTION PROFILES
# ==========================================

def split_double_lines(crop_img, logs):
    """
    Analyzes a crop to see if it accidentally contains TWO lines of text.
    Includes 'Descender Protection' to prevent cutting off tails (y, g, p).
    """
    # 1. Binarize
    gray = cv2.cvtColor(crop_img, cv2.COLOR_RGB2GRAY)
    _, thresh = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)

    # 2. Horizontal Projection
    h_proj = np.sum(thresh, axis=1)
    
    # Normalize
    max_val = np.max(h_proj)
    if max_val == 0: return [crop_img]
    h_proj = h_proj / max_val
    
    # 3. Find Peaks
    # Increased distance to 25 to prevent finding peaks within the same letter height
    peaks, _ = find_peaks(h_proj, height=0.2, distance=25)
    
    if len(peaks) < 2:
        return [crop_img]
        
    # 4. Analyze the Valley
    p1, p2 = peaks[0], peaks[1]
    valley_region = h_proj[p1:p2]
    
    if len(valley_region) == 0: return [crop_img]
    
    min_val = np.min(valley_region)
    min_idx = np.argmin(valley_region) + p1
    
    # ==========================
    # 🛡️ SAFETY GUARDRAILS
    # ==========================
    
    # CHECK 1: Valley Depth
    # Handwriting lines often touch. If the valley isn't DEEP (very low ink), don't split.
    # We lowered the threshold to 0.15 (15% ink density)
    if min_val > 0.15:
        return [crop_img]

    # CHECK 2: Edge Protection (The "Descender" Check)
    # If the split point is too close to the bottom (e.g., > 75% down the image),
    # it's almost certainly chopping off tails (y, g, p, q), not separating a new line.
    total_height = crop_img.shape[0]
    split_ratio = min_idx / total_height
    
    if split_ratio < 0.20 or split_ratio > 0.75:
        logs.append(f"   -> ⚠️ Refinement: Prevented split at {int(split_ratio*100)}% (Likely descenders)")
        return [crop_img]

    # If we pass checks, perform the split
    logs.append(f"   -> ✂️ Refinement: Split double line at Y={min_idx}")
    top_crop = crop_img[0:min_idx, :]
    bot_crop = crop_img[min_idx:, :]
    
    return [top_crop, bot_crop]
# ==========================================
# ⛓️ PIPELINE STEP: MERGING & ORDERING
# ==========================================

# ==========================================
# ⛓️ PIPELINE STEP: MERGING & ORDERING
# ==========================================

def smart_line_merger(raw_boxes, logs):
    """
    Groups words into lines using Centroid Clustering & Vertical Overlap.
    """
    # FIX: Use len() check because raw_boxes is a NumPy array, not a list
    if raw_boxes is None or len(raw_boxes) == 0:
        return []

    # 1. Clean & Format
    rects = []
    for box in raw_boxes:
        box = np.array(box).astype(np.float32)
        x1, y1 = np.min(box[:, 0]), np.min(box[:, 1])
        x2, y2 = np.max(box[:, 0]), np.max(box[:, 1])
        rects.append([x1, y1, x2, y2])
    
    rects = filter_nested_boxes(rects)
    logs.append(f"Valid Word Boxes: {len(rects)}")

    # 2. Sort by Y-Center (approximate top-down)
    rects.sort(key=lambda r: (r[1] + r[3]) / 2)
    
    lines = []
    
    while rects:
        # Start new line with the highest remaining box
        curr_line = [rects.pop(0)]
        
        # Find all other boxes that belong to this line
        remaining = []
        for r in rects:
            # Check overlap against the *average* vertical span of the current line
            overlap = get_vertical_overlap_ratio(curr_line[0], r)
            
            # 0.4 means they share 40% of their vertical height
            if overlap > 0.4: 
                curr_line.append(r)
            else:
                remaining.append(r)
        
        rects = remaining
        
        # Sort the collected line horizontally (Left to Right)
        curr_line.sort(key=lambda r: r[0])
        
        # Merge coordinates
        lx1 = min(r[0] for r in curr_line)
        ly1 = min(r[1] for r in curr_line)
        lx2 = max(r[2] for r in curr_line)
        ly2 = max(r[3] for r in curr_line)
        
        lines.append([lx1, ly1, lx2, ly2])

    # Final Sort of Lines (Top to Bottom)
    lines.sort(key=lambda r: r[1])
    return lines
# ==========================================
# 🚀 MAIN EXECUTION
# ==========================================

def process_handwriting(image):
    logs = ["--- STARTING PIPELINE ---"]
    
    if image is None: return None, [], "Please upload an image.", "Error"
    
    # 1. PRE-PROCESS
    # Convert to RGB array
    orig_np = np.array(image.convert("RGB"))
    
    # 2. DETECT (PaddleOCR)
    try:
        dt_boxes, _ = detector.text_detector(orig_np)
        if dt_boxes is None: dt_boxes = []
    except Exception as e:
        return image, [], f"Detector Failed: {e}", "\n".join(logs)

    if len(dt_boxes) == 0:
        return image, [], "No text detected.", "Logs end."

    # 3. MERGE WORDS -> LINES
    line_boxes = smart_line_merger(dt_boxes, logs)
    logs.append(f"Merged into {len(line_boxes)} lines.")

    # 4. RECOGNITION + REFINEMENT LOOP
    annotated_img = orig_np.copy()
    final_text_lines = []
    gallery_crops = []
    
    # Padding for crops (gives TrOCR context)
    PAD = 8 
    h_img, w_img, _ = orig_np.shape

    for i, box in enumerate(line_boxes):
        x1, y1, x2, y2 = map(int, box)
        
        # Add padding safely
        x1 = max(0, x1 - PAD); y1 = max(0, y1 - PAD)
        x2 = min(w_img, x2 + PAD); y2 = min(h_img, y2 + PAD)
        
        # Crop
        line_crop = orig_np[y1:y2, x1:x2]
        
        # --- REFINEMENT LOOP ---
        # Check if we accidentally merged two lines
        sub_crops = split_double_lines(line_crop, logs)
        
        for sub_crop in sub_crops:
            if sub_crop.shape[0] < 10 or sub_crop.shape[1] < 10: continue

            # Convert for TrOCR
            pil_crop = Image.fromarray(sub_crop)
            gallery_crops.append(pil_crop)
            
            # Inference
            with torch.no_grad():
                pixel_values = processor(images=pil_crop, return_tensors="pt").pixel_values.to(DEVICE)
                generated_ids = model.generate(pixel_values)
                text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
                
            if text.strip():
                final_text_lines.append(text)
        
        # Visualization (Draw the *original* merged box in Green)
        cv2.rectangle(annotated_img, (x1, y1), (x2, y2), (0, 200, 0), 2)
        cv2.putText(annotated_img, str(i+1), (x1, y1-5), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0,200,0), 1)

    full_text = "\n".join(final_text_lines)
    logs.append("--- PROCESSING COMPLETE ---")
    
    return Image.fromarray(annotated_img), gallery_crops, full_text, "\n".join(logs)


# ==========================================
# 🖥️ GRADIO INTERFACE
# ==========================================
css = """
#gallery { height: 300px; overflow-y: scroll; }
"""

with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
    gr.Markdown("## 📝 Scientific Handwriting OCR (Line-Level Refinement)")
    gr.Markdown("Uses PaddleOCR for detection, Geometry for merging, Projection Profiles for refinement, and TrOCR for reading.")
    
    with gr.Row():
        with gr.Column(scale=1):
            input_img = gr.Image(type="pil", label="Input Document")
            run_btn = gr.Button("Analyze & Transcribe", variant="primary")
        
        with gr.Column(scale=1):
            with gr.Tabs():
                with gr.Tab("Transcribed Text"):
                    output_txt = gr.Textbox(label="Result", lines=15, show_copy_button=True)
                with gr.Tab("Segmentation Map"):
                    output_img = gr.Image(label="Line Detection Map")
                with gr.Tab("System Logs"):
                    log_output = gr.Textbox(label="Process Logs", lines=15)

    gr.Markdown("### Line Segments (Input for TrOCR)")
    gallery = gr.Gallery(label="Refined Crops", columns=4, elem_id="gallery")

    run_btn.click(
        process_handwriting, 
        input_img, 
        [output_img, gallery, output_txt, log_output]
    )

if __name__ == "__main__":
    demo.launch()