This dataset is a high-resolution ship image dataset containing 8,359 images across 20 different ship categories. The dataset has undergone a meticulously designed 9-step data preprocessing pipeline, from initial collection to final classification and archiving, with each step carefully crafted to ensure data quality and consistency.

📊 Attribute	📝 Description
🖼️ Total Images	8,359
🏷️ Number of Categories	20
📐 Image Size	512 × 512 px
🎨 Image Format	PNG (RGB mode)
📦 Annotation Format	JSON (Natural language requests + Multilingual descriptions (including standardized descriptions) + Structured slots)

🗂️ Directory Structure

Dataset_ShipImages_PNG_px512(8359)_20_Type/
├── 📁 aircraft_carrier/       # Aircraft Carrier
│   ├── 📁 images/             # PNG image files
│   └── 📁 json/               # JSON annotation files
├── 📁 bulk_carrier/           # Bulk Carrier
│   ├── 📁 images/
│   └── 📁 json/
├── 📁 car_carrier/            # Car Carrier
├── 📁 container_ship/         # Container Ship
├── 📁 cruise_ship/            # Cruise Ship
├── 📁 destroyer/              # Destroyer
├── 📁 dredger/                # Dredger
├── 📁 fishing_vessel/         # Fishing Vessel
├── 📁 frigate/                # Frigate
├── 📁 offshore_support_vessel/# Offshore Support Vessel
├── 📁 oil_tanker/             # Oil Tanker
├── 📁 passenger_ferry/        # Passenger Ferry
├── 📁 patrol_boat/            # Patrol Boat
├── 📁 pilot_boat/             # Pilot Boat
├── 📁 recreational/           # Recreational Yacht
├── 📁 research_vessel/        # Research Vessel
├── 📁 rescue_vessel/          # Rescue Vessel
├── 📁 sailboat/               # Sailboat
├── 📁 submarine/              # Submarine
├── 📁 tugboat/                # Tugboat
└── 📄 README.md

Each category folder contains:

📁 images/ — All PNG image files for that category
📁 json/ — JSON intelligent annotation files corresponding one-to-one with images

🏷️ Category Distribution

Rank	🚢 Ship Category	🇨🇳 Chinese Name	📊 Count
1	`aircraft_carrier`	航空母舰	531
2	`sailboat`	帆船	529
3	`container_ship`	集装箱船	516
4	`cruise_ship`	邮轮	492
5	`passenger_ferry`	客运渡轮	476
6	`recreational`	休闲游艇	470
7	`bulk_carrier`	散货船	446
8	`car_carrier`	汽车运输船	427
9	`rescue_vessel`	救援船	415
10	`tugboat`	拖船	413
11	`submarine`	潜水艇	397
12	`research_vessel`	科考船	395
13	`offshore_support_vessel`	海洋工程支援船	382
14	`patrol_boat`	巡逻艇	378
15	`frigate`	护卫舰	374
16	`destroyer`	驱逐舰	359
17	`dredger`	疏浚船	350
18	`pilot_boat`	引航船	342
19	`fishing_vessel`	渔船	334
20	`oil_tanker`	油轮	333

📈 Classification Success Rate: 100.00% — All 8,359 images successfully classified, 0 failures

🔧 Data Preprocessing Pipeline

The dataset has undergone the following 9-step refined preprocessing pipeline to ensure the final output meets high-quality standards:

Step 1 📁 File Organization

Script: 00/01_file_organizer_*.py

Recursively scans and consolidates all files scattered across different directories into a single directory for unified management.

Step 2 🔄 Format Conversion

Script: 02_image_converter.py

Unifies image formats by converting all images to RGB mode, eliminating format differences from various sources.

Step 3 🧹 Quality Cleaning

Script: 03_image_cleaner.py

Removes low-resolution images and filters out samples that do not meet quality requirements, ensuring a visual quality baseline for the dataset.

Step 4 🔍 Sharpness Evaluation

Script: 05_image_sharpness_evaluator.py

Evaluates image sharpness, classifies images as sharp/blurry, and further screens high-quality samples.

Step 5 🤖 Intelligent Deduplication

Script: 06_image_deduplicator_optimized.py

Employs DINOv2 + FAISS for semantic-level deduplication, going beyond traditional perceptual hashing methods to identify and remove duplicate or highly similar images at the semantic level.

Step 6 📏 Resolution Adjustment

Script: 04_image_resizer.py

Unifies image dimensions to 512 × 512 px, ensuring consistent resolution across the dataset for subsequent model training.

Step 7 ✨ Super-Resolution Enhancement

Script: RealESRGAN/image_upscaler.py

Uses the RealESRGAN super-resolution model to enhance image quality, improve detail representation, and make images sharper and clearer.

Step 8 📦 Data Annotation

System: WenYao-Maritime_LLM_Annotation

An intelligent annotation system for high-resolution ship images based on Maritime Expert Knowledge Base and Large Language Model (LLM):

🧠 Utilizes RAG (Retrieval-Augmented Generation) technology combined with maritime domain expertise for precise annotation
⚡ Multi-threaded concurrent processing for efficient annotation
📝 Each image is annotated with: Chinese raw description, English standardized description, and structured feature slots

Step 9 📦 Dataset Splitting

Script: 07_Ship_image_classifier.py

Classifies and archives images by ship category, then uses stratified random sampling to split the complete dataset into a training set (80%), validation set (10%), and test set (10%). Stratification preserves the relative distribution of all 20 ship categories across the three subsets. Image files and their corresponding JSON annotations remain paired during splitting.

The resulting subsets are organized as follows:

dataset/
├── train/      # 80% of the data
├── val/        # 10% of the data
└── test/       # 10% of the data

📝 Annotation Format

Each image corresponds to a JSON annotation file containing the following fields:

{
    "subject": "High-resolution ship image dataset",
    "instance_name": "Ship_Image_xxxxx_XXXXX.PNG",
    "raw_description": {
        "language": "zh",
        "text": "Detailed Chinese description...",
        "natural_language_request": "Short Chinese generation prompt..."
    },
    "normalized_prompt": {
        "language": "en",
        "text": "English detailed description...",
        "natural_language_request": "English short generation prompt..."
    },
    "structured_slots": {
        "ship_type": "dredger",
        "function": "dredging",
        "view_angle": "rear_side_view",
        "scene": "coastal_water",
        "structure": ["bridge", "funnel", "crane", "mast", "superstructure", "pipeline"],
        "key_features": ["blue_hull", "white_superstructure", "discharging_sediment", "orange_cranes", "white_wake"],
        "style": "real_world_photography"
    }
}

Field Descriptions

Field	Description
`subject`	Dataset subject description
`instance_name`	Image file name
`raw_description`	Raw Chinese description: includes detailed text and natural language generation prompt
`normalized_prompt`	Standardized English description: includes detailed text and natural language generation prompt
`structured_slots.ship_type`	Ship type
`structured_slots.function`	Ship function
`structured_slots.view_angle`	Camera view angle
`structured_slots.scene`	Scene type
`structured_slots.structure`	List of visible ship structural components
`structured_slots.key_features`	List of key visual features
`structured_slots.style`	Image style

🛠️ Tech Stack

🔧 Technology	📌 Purpose
🤖 DINOv2	Semantic feature extraction for intelligent deduplication
⚡ FAISS	Efficient vector retrieval to support large-scale deduplication
✨ RealESRGAN	Super-resolution enhancement for improved image quality
🧠 LLM + RAG	Intelligent annotation based on maritime expert knowledge base
🐍 Python	Full pipeline script development

📊 Dataset Statistics Overview

============================================================
📊 Processing Summary
============================================================

📈 Overall Statistics:
   📁 Total Files: 8,359
   ✅ Successfully Classified: 8,359
   ❌ Classification Failures: 0
   📊 Success Rate: 100.00%

📂 File Operation Statistics:
   🖼️  Images Copied Successfully: 8,359
   🖼️  Image Copy Failures: 0
   📄 JSON Copied Successfully: 8,359
   📄 JSON Copy Failures: 0
============================================================

💡 Use Cases

🎯 Ship Image Classification — 20-class fine-grained ship classification tasks
🧠 Object Detection — Ship object detection and recognition
📝 Image Captioning — Multilingual image description tasks
🎨 Text-to-Image Generation — Prompt-based ship image generation
🔍 Retrieval-Augmented Search — Ship image semantic retrieval
📚 Maritime Research — Ship visual feature analysis research

📜 License

This dataset is intended for academic research and educational purposes only. Please comply with applicable laws and regulations and respect intellectual property rights.

🔨 Ethics and Responsible Use

This dataset is intended solely for academic research and educational purposes. Images have been reviewed to remove or avoid inclusion of identifiable human faces and sensitive personal information. Users must not use this dataset for surveillance, tracking, military targeting, or other harmful applications. If you believe any image infringes on copyright, privacy, or other rights, please contact the maintainer for removal.

Downloads last month: 13

Total file size:

4.88 GB