examples/extract_preprocess_frames.py

import cv2
import numpy as np
import os
from PIL import Image, ExifTags
import random
import argparse
import time

# Increase OpenCV/FFMPEG frame read retry attempts for long or unstable video reads
os.environ["OPENCV_FFMPEG_READ_ATTEMPTS"] = "50000"

# Default values (can be overridden via command line)
video_path = "path/to/video.mp4"
output_dir = "outputs/preprocessed"

frame_interval = 60          # Extract every nth frame
zoom_level = 5.0             # 5x zoom -> crop 1/5 of original frame, then resize to 640x640
crops_per_frame = 10         # Number of random crops per extracted frame
manual_crop = False          # If True, use one fixed crop position instead of random crops
crop_x_center = 0.5          # Manual crop center X (relative: 0.0 to 1.0)
crop_y_center = 0.5          # Manual crop center Y (relative: 0.0 to 1.0)


def extract_and_preprocess_frames(
    video_path,
    output_dir,
    frame_interval=10,
    zoom_level=5.0,
    crops_per_frame=5,
    manual_crop=False,
    crop_x=0.5,
    crop_y=0.5,
    start_frame=0,
    end_frame=None,
    segment_id=0,
):
    """
    Extract frames from a video segment and generate zoomed crops.

    Workflow:
    1) Read video frames in a specified frame range [start_frame, end_frame)
    2) Keep every nth frame (frame_interval)
    3) Optional EXIF-based orientation correction (mostly useful for images, harmless for video frames)
    4) Crop a zoom window (manual or random position)
    5) Resize crop to 640x640
    6) Apply CLAHE contrast enhancement
    7) Save as JPG with frame and crop position encoded in filename

    Returns:
        saved_count (int): Number of extracted frames processed (not total crops)
        total_saved (int): Total number of crops saved
    """
    # Create base output directory
    os.makedirs(output_dir, exist_ok=True)

    # Open video
    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        raise ValueError(f"Could not open video file: {video_path}")

    # Video metadata
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    fps = cap.get(cv2.CAP_PROP_FPS)

    # Clamp end_frame to video length
    if end_frame is None or end_frame > total_frames:
        end_frame = total_frames

    print(f"Video loaded: {video_path}")
    print(f"Total frames: {total_frames}")
    print(f"FPS: {fps}")
    print(f"Processing segment {segment_id + 1}: frames {start_frame} to {end_frame}")
    print(f"Extracting every {frame_interval} frames")
    print(f"Zooming {zoom_level * 100}% and cropping to 640x640")
    print(f"Crops per frame: {crops_per_frame}")
    print(f"Manual crop: {manual_crop}, Position: ({crop_x}, {crop_y})")

    # CLAHE for local contrast enhancement (helps visibility in crack-like textures)
    clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))

    # Seek to start frame
    cap.set(cv2.CAP_PROP_POS_FRAMES, start_frame)
    count = start_frame
    saved_count = 0  # Counts processed frames (not crops)

    # Save segment outputs in separate subfolder
    segment_dir = os.path.join(output_dir, f"segment_{segment_id + 1}")
    os.makedirs(segment_dir, exist_ok=True)

    while count < end_frame:
        ret, frame = cap.read()
        if not ret:
            print(f"Error reading frame at position {count}. Breaking out of loop.")
            break

        # Process every nth frame inside this segment
        if (count - start_frame) % frame_interval == 0:
            # Convert to PIL for optional EXIF orientation correction
            pil_img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))

            # EXIF orientation correction (videos usually don't have EXIF, so this often does nothing)
            try:
                orientation_tag = None
                for tag_id, tag_name in ExifTags.TAGS.items():
                    if tag_name == "Orientation":
                        orientation_tag = tag_id
                        break

                exif = dict(pil_img.getexif().items())

                if orientation_tag is not None and orientation_tag in exif:
                    if exif[orientation_tag] == 3:
                        pil_img = pil_img.rotate(180, expand=True)
                    elif exif[orientation_tag] == 6:
                        pil_img = pil_img.rotate(270, expand=True)
                    elif exif[orientation_tag] == 8:
                        pil_img = pil_img.rotate(90, expand=True)
            except (AttributeError, KeyError, IndexError, TypeError):
                # No EXIF or EXIF not readable
                pass

            # Back to OpenCV BGR
            frame = cv2.cvtColor(np.array(pil_img), cv2.COLOR_RGB2BGR)

            # Original frame size
            height, width = frame.shape[:2]

            # Compute crop window size from zoom factor
            # Example: zoom=5 -> crop 1/5 width and 1/5 height, then upscale to 640x640
            crop_width = int(width / zoom_level)
            crop_height = int(height / zoom_level)

            # Safety check for invalid crop sizes
            if crop_width <= 0 or crop_height <= 0:
                print(f"Skipping frame {count}: invalid crop size ({crop_width}, {crop_height})")
                count += 1
                continue

            if manual_crop:
                # Single deterministic crop at user-defined relative position
                process_crop(
                    frame, count, 0, crop_x, crop_y,
                    crop_width, crop_height, width, height,
                    segment_dir, clahe
                )
                saved_count += 1
            else:
                # Multiple random crops per frame
                for i in range(crops_per_frame):
                    # Avoid edges slightly to reduce out-of-frame crop clipping
                    random_x = random.uniform(0.1, 0.9)
                    random_y = random.uniform(0.1, 0.9)

                    process_crop(
                        frame, count, i, random_x, random_y,
                        crop_width, crop_height, width, height,
                        segment_dir, clahe
                    )
                saved_count += 1

            if saved_count % 10 == 0:
                print(f"Segment {segment_id + 1}: Processed {saved_count} extracted frames (current frame={count})")

        count += 1

    cap.release()

    # Total number of crop images written
    total_saved = saved_count * (1 if manual_crop else crops_per_frame)

    print(
        f"Segment {segment_id + 1} completed! "
        f"Processed {saved_count} extracted frames and saved {total_saved} crops."
    )
    return saved_count, total_saved


def process_crop(
    frame,
    frame_count,
    crop_index,
    rel_x,
    rel_y,
    crop_width,
    crop_height,
    width,
    height,
    output_dir,
    clahe,
):
    """
    Create one crop from a frame, resize to 640x640, enhance contrast, and save.

    Args:
        rel_x, rel_y: Relative crop center coordinates in [0, 1]
    """
    # Convert relative center coords to pixel coordinates
    center_x = int(rel_x * width)
    center_y = int(rel_y * height)

    # Top-left crop corner
    start_x = center_x - (crop_width // 2)
    start_y = center_y - (crop_height // 2)

    # Clamp crop start so crop stays inside frame
    start_x = max(0, min(start_x, width - crop_width))
    start_y = max(0, min(start_y, height - crop_height))

    # Compute crop end
    end_x = min(start_x + crop_width, width)
    end_y = min(start_y + crop_height, height)

    # Final adjustment if crop was clipped
    if end_x - start_x < crop_width:
        start_x = max(0, end_x - crop_width)
    if end_y - start_y < crop_height:
        start_y = max(0, end_y - crop_height)

    try:
        # Crop region from original frame
        cropped = frame[start_y:end_y, start_x:end_x]

        # Skip empty crops (rare edge case)
        if cropped.size == 0:
            print(f"Empty crop at frame {frame_count}, crop {crop_index}")
            return

        # Resize to model-friendly input size
        zoomed = cv2.resize(cropped, (640, 640), interpolation=cv2.INTER_LINEAR)

        # Apply CLAHE
        if len(zoomed.shape) == 3:
            # Color image: apply CLAHE channel-wise in BGR space
            # Note: This may shift colors slightly. For more natural results, apply on LAB L-channel.
            enhanced = cv2.merge([
                clahe.apply(zoomed[:, :, 0]),
                clahe.apply(zoomed[:, :, 1]),
                clahe.apply(zoomed[:, :, 2]),
            ])
        else:
            # Grayscale image
            enhanced = clahe.apply(zoomed)

        # Filename convention
        # crop_index == 0 is first crop; in manual mode this is the only crop
        if crop_index == 0:
            filename = f"frame5_{frame_count:06d}"
        else:
            filename = f"frame5_{frame_count:06d}_crop{crop_index}"

        # Encode crop center position (% of frame) for traceability
        pos_info = f"_x{int(rel_x * 100):03d}_y{int(rel_y * 100):03d}"
        frame_filename = os.path.join(output_dir, f"{filename}{pos_info}.jpg")

        cv2.imwrite(frame_filename, enhanced)

    except Exception as e:
        print(f"Error processing crop at frame {frame_count}, crop {crop_index}: {e}")


def process_video_in_segments(
    video_path,
    output_dir,
    frame_interval,
    zoom_level,
    crops_per_frame,
    manual_crop,
    crop_x,
    crop_y,
    segment_size=5000,
    overlap=100,
):
    """
    Process video in segments to avoid memory/decoder instability on long videos.

    Note:
    - Overlap can help avoid missing frames near segment boundaries.
    - But overlap can also create duplicate outputs if the same frame is processed in two segments.
    """
    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        raise ValueError(f"Could not open video file: {video_path}")

    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    cap.release()

    print(f"Total frames in video: {total_frames}")
    print(f"Processing in segments of {segment_size} frames with {overlap} frame overlap")

    # Prevent invalid step size
    if segment_size <= overlap:
        raise ValueError("segment_size must be greater than overlap")

    # Segment start indices
    start_frames = list(range(0, total_frames, segment_size - overlap))

    total_frames_processed = 0  # processed extracted frames
    total_crops_processed = 0   # saved crop images

    for i, start_frame in enumerate(start_frames):
        end_frame = min(start_frame + segment_size, total_frames)

        print(f"\n{'=' * 80}")
        print(f"Processing segment {i + 1}/{len(start_frames)}: frames {start_frame} to {end_frame}")
        print(f"{'=' * 80}\n")

        # Small pause between segments (helps file handles / decoder stability)
        if i > 0:
            time.sleep(2)

        try:
            frames_processed, crops_processed = extract_and_preprocess_frames(
                video_path=video_path,
                output_dir=output_dir,
                frame_interval=frame_interval,
                zoom_level=zoom_level,
                crops_per_frame=crops_per_frame,
                manual_crop=manual_crop,
                crop_x=crop_x,
                crop_y=crop_y,
                start_frame=start_frame,
                end_frame=end_frame,
                segment_id=i,
            )

            total_frames_processed += frames_processed
            total_crops_processed += crops_processed

        except Exception as e:
            print(f"Error processing segment {i + 1}: {e}")
            print("Continuing with next segment...")

    print(f"\n{'=' * 80}")
    print(
        f"Processing complete! "
        f"Processed {total_frames_processed} extracted frames and saved {total_crops_processed} crops."
    )
    print(f"{'=' * 80}")


if __name__ == "__main__":
    parser = argparse.ArgumentParser(
        description="Extract frames from a video and generate zoomed crops with optional CLAHE enhancement"
    )

    # Input/output
    parser.add_argument("--video", type=str, default=video_path, help="Path to input video")
    parser.add_argument("--output", type=str, default=output_dir, help="Directory to save processed crops")

    # Extraction/cropping settings
    parser.add_argument("--interval", type=int, default=frame_interval, help="Extract every nth frame")
    parser.add_argument("--zoom", type=float, default=zoom_level, help="Zoom factor (e.g., 5.0 = 500%%)")
    parser.add_argument("--crops", type=int, default=crops_per_frame, help="Random crops per extracted frame")
    parser.add_argument("--manual", action="store_true", help="Use one manual crop position instead of random crops")
    parser.add_argument("--crop_x", type=float, default=crop_x_center, help="Manual crop center X in [0,1]")
    parser.add_argument("--crop_y", type=float, default=crop_y_center, help="Manual crop center Y in [0,1]")

    # Segmentation settings (for long videos)
    parser.add_argument("--segment_size", type=int, default=5000, help="Frames per segment")
    parser.add_argument("--overlap", type=int, default=100, help="Segment overlap in frames")

    args = parser.parse_args()

    # Basic input validation
    if not args.video or not os.path.isfile(args.video):
        print(f"Error: Video file '{args.video}' does not exist.")
        raise SystemExit(1)

    if args.interval <= 0:
        print("Error: --interval must be > 0")
        raise SystemExit(1)

    if args.zoom <= 0:
        print("Error: --zoom must be > 0")
        raise SystemExit(1)

    if args.crops <= 0 and not args.manual:
        print("Error: --crops must be > 0 when not using --manual")
        raise SystemExit(1)

    if not (0.0 <= args.crop_x <= 1.0 and 0.0 <= args.crop_y <= 1.0):
        print("Error: --crop_x and --crop_y must be in [0,1]")
        raise SystemExit(1)

    try:
        process_video_in_segments(
            video_path=args.video,
            output_dir=args.output,
            frame_interval=args.interval,
            zoom_level=args.zoom,
            crops_per_frame=args.crops,
            manual_crop=args.manual,
            crop_x=args.crop_x,
            crop_y=args.crop_y,
            segment_size=args.segment_size,
            overlap=args.overlap,
        )
    except Exception as e:
        print(f"An error occurred: {e}")