Python | OpenCV 專案：天堂私服 YOLOv8 物件偵測實戰

# collect_screenshots.py
import mss
import cv2
import numpy as np
import os
import time
import pygetwindow as gw

SAVE_DIR  = "dataset/raw_images"
INTERVAL  = 0.5     # 每隔 0.5 秒截一張
TARGET    = 500     # 目標張數
WINDOW_TITLE = "Lineage"    # 遊戲視窗標題（部分符合即可）

os.makedirs(SAVE_DIR, exist_ok=True)

# 找到遊戲視窗位置
def get_game_window():
    wins = [w for w in gw.getAllWindows() if WINDOW_TITLE in w.title]
    if not wins:
        print(f"找不到視窗：{WINDOW_TITLE}")
        return None
    w = wins[0]
    return {"left": w.left, "top": w.top,
            "width": w.width, "height": w.height}

region = get_game_window()
if region is None:
    # 找不到視窗時，截全螢幕
    region = {"left": 0, "top": 0, "width": 1920, "height": 1080}

count = 0
print(f"開始蒐集，目標 {TARGET} 張，間隔 {INTERVAL} 秒")
print("切換到遊戲視窗後自動開始截圖，Ctrl+C 停止")

with mss.mss() as sct:
    while count < TARGET:
        screenshot = sct.grab(region)
        img = np.array(screenshot)
        img = cv2.cvtColor(img, cv2.COLOR_BGRA2BGR)

        fname = os.path.join(SAVE_DIR, f"{count:05d}.jpg")
        cv2.imwrite(fname, img)
        count += 1

        if count % 50 == 0:
            print(f"已蒐集：{count}/{TARGET}")

        time.sleep(INTERVAL)

print(f"蒐集完成！共 {count} 張，儲存於 {SAVE_DIR}")

# split_dataset.py
import os
import shutil
import random

SRC_IMAGES = "dataset/raw_images"
SRC_LABELS = "dataset/raw_labels"
DST_ROOT   = "dataset"
VAL_RATIO  = 0.2
random.seed(42)

for split in ["train", "val"]:
    os.makedirs(f"{DST_ROOT}/images/{split}", exist_ok=True)
    os.makedirs(f"{DST_ROOT}/labels/{split}", exist_ok=True)

files = [f for f in os.listdir(SRC_IMAGES)
         if f.lower().endswith((".jpg", ".png"))]
random.shuffle(files)

val_set = set(files[:int(len(files) * VAL_RATIO)])

for fname in files:
    split = "val" if fname in val_set else "train"
    stem  = os.path.splitext(fname)[0]

    shutil.copy(f"{SRC_IMAGES}/{fname}",
                f"{DST_ROOT}/images/{split}/{fname}")

    lbl = f"{SRC_LABELS}/{stem}.txt"
    if os.path.exists(lbl):
        shutil.copy(lbl, f"{DST_ROOT}/labels/{split}/{stem}.txt")
    else:
        open(f"{DST_ROOT}/labels/{split}/{stem}.txt", "w").close()

print(f"train: {len(files)-len(val_set)}, val: {len(val_set)}")

# create_yaml.py
import os

# 從 LabelImg 產生的 classes.txt 讀取類別
with open("dataset/raw_labels/classes.txt") as f:
    classes = [line.strip() for line in f if line.strip()]

yaml_content = f"""path: ./dataset
train: images/train
val:   images/val

nc: {len(classes)}
names:
"""
for i, cls in enumerate(classes):
    yaml_content += f"  {i}: {cls}\n"

with open("data.yaml", "w") as f:
    f.write(yaml_content)

print("data.yaml 已產生：")
print(yaml_content)

# train.py
from ultralytics import YOLO
import torch

print(f"GPU：{'可用 - ' + torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'CPU'}")

model = YOLO("yolov8s.pt")

model.train(
    data="data.yaml",
    epochs=100,
    imgsz=640,
    batch=16,
    device=0 if torch.cuda.is_available() else "cpu",
    workers=0,          # Windows 必設 0
    patience=20,
    name="lineage_detector",   # 不顯式指定 project，避免不同版本加多一層 runs/detect
    plots=True,
)

# 訓練完成後印出實際儲存路徑（不同版本路徑層級可能不同）
print(f"\n訓練完成！實際模型路徑：{model.trainer.best}")

# detect_realtime.py
from ultralytics import YOLO
from pathlib import Path
import mss
import cv2
import numpy as np
import pygetwindow as gw

CONF         = 0.5
WINDOW_TITLE = "Lineage"

def find_latest_model(search_root="runs"):
    """自動找出 runs/ 目錄下最近訓練的 best.pt

    Ultralytics 在不同版本下實際存檔路徑可能是：
      runs/detect/lineage_detector/weights/best.pt
      runs/detect/runs/detect/lineage_detector/weights/best.pt
      runs/detect/lineage_detector2/weights/best.pt   （訓練多次時）
    用 rglob 通吃所有情況，並依修改時間挑最新。
    """
    candidates = list(Path(search_root).rglob("best.pt"))
    if not candidates:
        return None
    return str(max(candidates, key=lambda p: p.stat().st_mtime))

MODEL_PATH = find_latest_model()
if MODEL_PATH is None:
    print("找不到任何 best.pt，請先執行 train.py 完成訓練")
    exit()
print(f"載入模型：{MODEL_PATH}")

model = YOLO(MODEL_PATH)

def find_game_window(title_keyword):
    """回傳符合標題關鍵字，且可見、未最小化、尺寸正常的視窗"""
    for w in gw.getAllWindows():
        if (title_keyword.lower() in w.title.lower()
                and w.visible
                and not w.isMinimized
                and w.width > 0 and w.height > 0):
            return w
    return None

def window_to_region(win):
    return {"left": win.left, "top": win.top,
            "width": win.width, "height": win.height}

# 只抓 Lineage 視窗（找不到就列出所有可見視窗協助診斷，不再 fallback 成全螢幕）
game_win = find_game_window(WINDOW_TITLE)
if game_win is None:
    print(f"[錯誤] 找不到可見的『{WINDOW_TITLE}』視窗。目前可見視窗：")
    for w in gw.getAllWindows():
        if w.title.strip() and w.visible:
            print(f"  - {w.title}")
    print("\n請確認遊戲已開啟且未最小化，或調整 WINDOW_TITLE 讓它能匹配遊戲視窗標題。")
    exit()

print(f"目標視窗：{game_win.title}  位置：({game_win.left},{game_win.top})  大小：{game_win.width}x{game_win.height}")
print("即時偵測啟動，按 q 離開（需點選偵測視窗）")

# 開始迴圈前，明確建立「單一」顯示視窗，尺寸與遊戲視窗一致
DISPLAY_NAME = "Lineage Detector"
cv2.namedWindow(DISPLAY_NAME, cv2.WINDOW_NORMAL)
cv2.resizeWindow(DISPLAY_NAME, game_win.width, game_win.height)

REFRESH_EVERY = 30   # 每 N 幀重新查詢視窗位置，遊戲視窗被拖移也能跟上
region = window_to_region(game_win)
frame_count = 0

try:
    with mss.mss() as sct:
        while True:
            # 定期更新視窗位置，並偵測遊戲視窗是否被關閉或最小化
            if frame_count % REFRESH_EVERY == 0:
                game_win = find_game_window(WINDOW_TITLE)
                if game_win is None:
                    print("偵測到遊戲視窗消失或被最小化，結束偵測")
                    break
                region = window_to_region(game_win)
            frame_count += 1

            screenshot = sct.grab(region)
            frame = cv2.cvtColor(np.array(screenshot), cv2.COLOR_BGRA2BGR)

            results = model(frame, conf=CONF, verbose=False)
            boxes   = results[0].boxes

            # 手動畫框（不呼叫 results.plot()，避免某些版本會開額外視窗或觸發內建 show）
            for box in boxes:
                cls_id = int(box.cls)
                conf   = float(box.conf)
                name   = model.names[cls_id]
                x1, y1, x2, y2 = map(int, box.xyxy[0])

                cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
                cv2.putText(frame, f"{name} {conf:.2f}", (x1, max(y1 - 8, 15)),
                            cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)

            cv2.imshow(DISPLAY_NAME, frame)
            if cv2.waitKey(1) & 0xFF == ord("q"):
                break
finally:
    cv2.destroyAllWindows()