遙感影像各類油罐檢測共分為外浮頂油罐封閉頂油罐球形壓力罐水塔沉淀罐五種類型超過26000張影像采用voc格式標注512×215尺寸2GB使用YOLOv8來訓練一個包含超過26000張遙感影像的油罐檢測數(shù)據(jù)集。這個數(shù)據(jù)集包含5個類別已標注為VOC格式可以直接用于模型訓練。數(shù)據(jù)集描述數(shù)據(jù)量超過26000張影像類別0: 外浮頂油罐External Floating Roof Tank1: 封閉頂油罐Fixed Roof Tank2: 球形壓力罐Spherical Pressure Tank3: 水塔Water Tower4: 沉淀罐Settling Tank標注格式VOC格式圖像尺寸512×215數(shù)據(jù)大小約2GB應用場景遙感影像油罐檢測數(shù)據(jù)集組織假設(shè)你的數(shù)據(jù)集目錄結(jié)構(gòu)如下深色版本oil_tank_dataset/├── images/│ ├── 000001.jpg│ ├── 000002.jpg│ └── …├── annotations/│ ├── 000001.xml│ ├── 000002.xml│ └── …└── data.yaml # 數(shù)據(jù)配置文件數(shù)據(jù)配置文件創(chuàng)建或確認data.yaml文件是否正確配置了數(shù)據(jù)集路徑和類別信息yaml深色版本train: ./images/train/ # 訓練集圖像路徑val: ./images/val/ # 驗證集圖像路徑test: ./images/test/ # 測試集圖像路徑Classesnc: 5 # 類別數(shù)量names:External Floating Roof TankFixed Roof TankSpherical Pressure TankWater TowerSettling Tank # 類別名稱列表數(shù)據(jù)集劃分將數(shù)據(jù)集劃分為訓練集、驗證集和測試集?？梢允褂靡韵履_本python深色版本import osimport randomfrom shutil import copyfile定義源目錄和目標目錄source_images_dir ‘./oil_tank_dataset/images’source_annotations_dir ‘./oil_tank_dataset/annotations’target_train_dir ‘./oil_tank_dataset/images/train’target_val_dir ‘./oil_tank_dataset/images/val’target_test_dir ‘./oil_tank_dataset/images/test’target_train_annotations_dir ‘./oil_tank_dataset/annotations/train’target_val_annotations_dir ‘./oil_tank_dataset/annotations/val’target_test_annotations_dir ‘./oil_tank_dataset/annotations/test’創(chuàng)建目標目錄os.makedirs(target_train_dir, exist_okTrue)os.makedirs(target_val_dir, exist_okTrue)os.makedirs(target_test_dir, exist_okTrue)os.makedirs(target_train_annotations_dir, exist_okTrue)os.makedirs(target_val_annotations_dir, exist_okTrue)os.makedirs(target_test_annotations_dir, exist_okTrue)獲取所有圖像文件all_images [f for f in os.listdir(source_images_dir) if f.endswith(‘.jpg’)]random.shuffle(all_images)劃分數(shù)據(jù)集train_ratio 0.8val_ratio 0.1test_ratio 0.1train_split int(train_ratio * len(all_images))val_split train_split int(val_ratio * len(all_images))train_images all_images[:train_split]val_images all_images[train_split:val_split]test_images all_images[val_split:]復制圖像文件和對應的標注文件def copy_files(image_list, target_image_dir, target_annotation_dir):for img in image_list:copyfile(os.path.join(source_images_dir, img), os.path.join(target_image_dir, img))annotation img.replace(‘.jpg’, ‘.xml’)copyfile(os.path.join(source_annotations_dir, annotation), os.path.join(target_annotation_dir, annotation))copy_files(train_images, target_train_dir, target_train_annotations_dir)copy_files(val_images, target_val_dir, target_val_annotations_dir)copy_files(test_images, target_test_dir, target_test_annotations_dir)轉(zhuǎn)換VOC標注為YOLO格式首先我們需要將VOC格式的標注文件轉(zhuǎn)換為YOLO格式?？梢允褂肞ython腳本來完成這個任務。python深色版本import osimport xml.etree.ElementTree as ETimport shutil定義類別映射class_map {‘External Floating Roof Tank’: 0,‘Fixed Roof Tank’: 1,‘Spherical Pressure Tank’: 2,‘Water Tower’: 3,‘Settling Tank’: 4}定義轉(zhuǎn)換函數(shù)def convert_voc_to_yolo(voc_file, yolo_file, image_size, class_map):tree ET.parse(voc_file)root tree.getroot()with open(yolo_file, w) as f: for obj in root.findall(object): class_name obj.find(name).text if class_name not in class_map: continue class_id class_map[class_name] bbox obj.find(bndbox) x_min int(bbox.find(xmin).text) y_min int(bbox.find(ymin).text) x_max int(bbox.find(xmax).text) y_max int(bbox.find(ymax).text) x_center (x_min x_max) / 2.0 / image_size[0] y_center (y_min y_max) / 2.0 / image_size[1] width (x_max - x_min) / image_size[0] height (y_max - y_min) / image_size[1] f.write(f{class_id} {x_center} {y_center} {width} {height}\n)讀取圖像尺寸def get_image_size(image_path):from PIL import Imagewith Image.open(image_path) as img:return img.width, img.height轉(zhuǎn)換所有標注文件def convert_all_annotations(image_dir, annotation_dir, output_dir):if not os.path.exists(output_dir):os.makedirs(output_dir)for filename in os.listdir(annotation_dir): if filename.endswith(.xml): image_filename filename.replace(.xml, .jpg) image_path os.path.join(image_dir, image_filename) voc_file os.path.join(annotation_dir, filename) yolo_file os.path.join(output_dir, filename.replace(.xml, .txt)) image_size get_image_size(image_path) convert_voc_to_yolo(voc_file, yolo_file, image_size, class_map)調(diào)用轉(zhuǎn)換函數(shù)image_dir ‘./oil_tank_dataset/images’annotation_dir ‘./oil_tank_dataset/annotations’output_dir ‘./oil_tank_dataset/labels’convert_all_annotations(image_dir, annotation_dir, output_dir)安裝YOLOv8如果你還沒有安裝YOLOv8可以使用以下命令安裝pip install ultralytics訓練模型使用YOLOv8訓練模型的命令非常簡單你可以直接使用以下命令開始訓練cd path/to/oil_tank_dataset/克隆YOLOv8倉庫git clone https://github.com/ultralytics/ultralytics.gitcd ultralytics開始訓練python yolo.py detect train data…/data.yaml modelyolov8n.pt epochs100 imgsz512 batch16在這個命令中data…/data.yaml指定數(shù)據(jù)配置文件。modelyolov8n.pt指定預訓練權(quán)重這里使用的是YOLOv8的小模型。epochs100訓練輪數(shù)。imgsz512輸入圖像的寬度高度為215但YOLOv8通常需要正方形輸入可以考慮調(diào)整圖像大小或使用其他方法處理。batch16批量大小。模型評估訓練完成后可以使用以下命令評估模型在驗證集上的表現(xiàn)python yolo.py detect val data…/data.yaml modelruns/detect/train/weights/best.pt imgsz512這里的runs/detect/train/weights/best.pt是訓練過程中產(chǎn)生的最佳模型權(quán)重文件。模型預測你可以使用訓練好的模型對新圖像進行預測python yolo.py detect predict sourcepath/to/your/image.jpg modelruns/detect/train/weights/best.pt imgsz512 conf0.4 iou0.5查看訓練結(jié)果訓練過程中的日志和結(jié)果會保存在runs/detect/目錄下你可以查看訓練過程中的損失、精度等信息。數(shù)據(jù)增強為了進一步提高模型性能可以使用數(shù)據(jù)增強技術(shù)。以下是一個簡單的數(shù)據(jù)增強示例安裝albumentations庫pip install -U albumentations在yolo.py中添加數(shù)據(jù)增強pimport albumentations as Afrom albumentations.pytorch import ToTensorV2import cv2定義數(shù)據(jù)增強transform A.Compose([A.RandomSizedCrop(min_max_height(400, 512), height512, width512, p0.5),A.HorizontalFlip(p0.5),A.VerticalFlip(p0.5),A.Rotate(limit10, p0.5, border_modecv2.BORDER_CONSTANT),A.ColorJitter(brightness0.2, contrast0.2, saturation0.2, hue0.2, p0.5),A.GaussNoise(var_limit(10.0, 50.0), p0.5),A.Normalize(mean[0.485, 0.456, 0.406], std[0.229, 0.224, 0.225]),ToTensorV2()], bbox_paramsA.BboxParams(format‘yolo’, label_fields[‘class_labels’]))在數(shù)據(jù)加載器中應用數(shù)據(jù)增強def collate_fn(batch):images, targets zip(*batch)transformed_images []transformed_targets []for img, target in zip(images, targets): bboxes target[bboxes] class_labels target[labels] augmented transform(imageimg, bboxesbboxes, class_labelsclass_labels) transformed_images.append(augmented[image]) transformed_targets.append({ bboxes: augmented[bboxes], labels: augmented[class_labels] }) return torch.stack(transformed_images), transformed_targets注意事項數(shù)據(jù)集質(zhì)量確保數(shù)據(jù)集的質(zhì)量包括清晰度、標注準確性等。模型選擇可以選擇更強大的模型版本如YOLOv8m、YOLOv8l等以提高性能。超參數(shù)調(diào)整根據(jù)實際情況調(diào)整超參數(shù)如批量大小batch、圖像大小imgsz等。監(jiān)控性能訓練過程中監(jiān)控損失函數(shù)和mAP指標確保模型收斂。通過上述步驟你可以使用YOLOv8來訓練一個遙感影像油罐檢測數(shù)據(jù)集并使用訓練好的模型進行預測。