2026.03 · Pebblous Data Communication Team

Reading time: ~10 min · 한국어

Executive Summary

This article is based on the analysis results from DataClinic Report #131. The Industrial Waste Image dataset is the largest industrial waste image collection provided by AI Hub (Korea's national AI data platform under the Ministry of Science and ICT). Despite its impressive scale of 72 waste categories and 1 million images, it scored only 51 (Poor) in the DataClinic diagnosis. Class imbalance reaches up to 3,978x, and the dataset is flooded with visually similar images.

51
DataClinic Overall Score
72
Waste Classes
1M
Total Images
3,978x
Max Class Imbalance

DataClinic Grade Summary

L1 IntegrityFair
L1 MissingFair
L1 Class BalancePoor
L1 StatisticsPoor
L2 DataLensNo Issues
L2 GeometryGood
L2 DistributionGood
L3 DataLensNo Issues
L3 GeometryGood
L3 DistributionGood
💡 How is the 51-point score calculated?

The DataClinic composite score is a weighted sum of sub-grades across L1 (basic quality), L2 (general neural network), and L3 (domain-specific), yielding a 0–100 score. This dataset scored 'Good' on L2/L3 distribution and geometry, but 'Bad' on L1 class balance and statistics — dragging down the overall score significantly. In short, the data structure is sound, but basic quality management failed — a textbook case.

📊 DataClinic's 3-Level Diagnostic Framework

DataClinic diagnoses datasets at 3 levels of depth. From surface-level statistics to domain-specific analysis, each level uncovers increasingly precise quality issues.

L1

Basic Quality Diagnostics

Checks missing values, class balance, resolution, and statistical diversity — the dataset's fundamental health. The fastest way to spot problems.

L2

DataLens Analysis (General-Purpose Neural Network)

Vectorizes images via Wolfram ImageIdentify Net V2 (1,280 dimensions), then analyzes geometric relationships and density distributions between classes. Seeing data through "AI's eyes."

L3

Domain-Specific Analysis (Optimized Lens)

Optimizes dimensions for the specific domain (136 dimensions for this dataset). Captures domain-specific patterns and outliers that general-purpose L2 misses.

Dataset Overview — Government-Built Waste AI Data

South Korea generates approximately 200 million tons of industrial waste annually, much of which is still sorted manually. Automating waste classification is a critical challenge for achieving carbon neutrality and a circular economy, with AI-based sorting systems at the center of the solution. Against this backdrop, the government embarked on building large-scale waste image datasets.

Since 2019, the Korean government has been building national AI training data through AI Hub. The Industrial Waste Image dataset is Korea's largest industrial waste image collection produced as part of this initiative. It consists of high-resolution images of 72 types of waste captured at actual factories and industrial sites.

Industrial waste image dataset full collage

▲ Industrial waste image dataset collage — 72 diverse waste types including Metal Waste, Textile Waste, Glass & Ceramic, Synthetic Resin, and more

🏭

National AI Data

Official AI Hub (MSIT) dataset
Commercial use allowed

📸

Real-World High-Res

Up to 3,024×4,032px
Shot at actual industrial sites

♻️

Recycling & Classification AI

Classification, detection, segmentation
Multi-purpose applications

Filename Structure Decoded: 01_X020_C056_0223_3.jpg
First 2 digits: major category number · X-number: shooting location · C-number: sub-class code · MMDD: shooting date · Last digit: sequential shot number of same subject

72 Waste Categories — What Kinds of Trash Are There?

The 72 classes are organized under 6 major categories with further sub-classifications. From metal cans to waste concrete sewer pipes, the dataset covers nearly every type of waste found at industrial sites.

🔩 Metal Waste (금속류)
Others Lacquer Can Paint Can Wire/Rebar Cans

Iron, aluminum, and metal containers. High recycling value. Lacquer and paint cans are managed separately as hazardous waste.

🏺 Glass & Ceramic (유리도자기류)
Ceramic Window Glass Others

Ceramic fragments, glass bottles, window glass, etc. The most 'typical' waste for AI — #1 high density (density 2.13).

🧵 Textile Waste (폐섬유)
Others Fabric Sleeping Bag Tent

Clothing and fabric waste. Camping waste like sleeping bags and tents are separately classified. Contains many L1 mean images.

🧴 Synthetic Resin (합성수지)
Plastic Vinyl/Film Styrofoam

Plastic and vinyl waste. The most difficult class for AI — #1 low density (density 0.36). Extremely high shape diversity.

📄 Paper Waste (폐지류)
Wallpaper Corrugated Cardboard Mixed Paper

Paper and cardboard waste. Wallpaper's diverse patterns cause AI recognition confusion. Frequently appears as low-density outliers.

🧱 Waste Concrete (폐콘크리트)
Sewer Pipe Concrete Debris Brick

Construction waste. Sewer pipes with their cylindrical structure are the most 'different' waste from window glass — appearing as dissimilar pairs.

Level 1 — Basic Quality Diagnosis

✅ Strengths

  • 🎨 RGB Channel Consistency: All images in unified RGB format
  • 📐 High Resolution: Min 1,920×1,080px to Max 3,024×4,032px
  • 🏷️ Label Integrity: No classification label consistency issues
  • Minimal Missing Data: Only 2 images missing out of 1M (0.0002%)

⚠️ Key Issues

  • 📊 Class Balance: Poor — Min 20 images vs Max 79,560 images
  • 🔄 Visual Diversity: Poor — Too many similar images
  • 📏 Resolution Inconsistency: Large gap between min and max resolution

Class Mean Images — The "Face" of Each Waste Through AI's Eyes

These are pixel-averaged results of each class's images. The blurrier the result, the higher the image diversity within that class; the sharper it looks, the more similar images are being repeated.

Metal-Others representative Actual
Metal-Others mean image Mean
Metal-Others
Metal-Lacquer representative Actual
Metal-Lacquer mean image Mean
Metal-Lacquer
Textile-Others representative Actual
Textile-Others mean image Mean
Textile-Others
Textile-Fabric representative Actual
Textile-Fabric mean image Mean
Textile-Fabric
Textile-Sleeping Bag representative Actual
Textile-Sleeping Bag mean image Mean
Textile-Sleeping Bag
Textile-Tent representative Actual
Textile-Tent mean image Mean
Textile-Tent

▲ Each card — Left: representative sample (actual image) / Right: mean image (pixel average of all images)

💡 Insight — Sharp Mean Images for Tent & Sleeping Bag: The mean images for Textile-Tent and Textile-Sleeping Bag are relatively sharp. This means images within these classes were shot with similar compositions and backgrounds — a cause of the L1 Statistics: Poor rating. Additional photography from diverse angles and environments is needed.

Class Imbalance Deep Dive — The Shock of 3,978x

The most critical issue with this dataset is class imbalance. While the average is 11,257 images per class, the standard deviation of 20,343 is 1.8 times larger than the mean, indicating extreme skewness.

⚠️ Class Imbalance Status

20
Min Class
11,257
Class Average
79,560
Max Class

Min class (20 images) vs Max class (79,560 images) → 3,978x difference

Imbalance Visualization

Max Class (e.g., Synthetic Resin-Vinyl)79,560
100%
Average Class11,257
14%
Min Class20
💡 Why Is Imbalance a Problem? AI models become biased toward classes with more data. The minimum class (20 images) becomes waste that AI has barely "seen." When AI encounters such waste in real-world settings, misclassification probability increases significantly. Data Bulk-up (reinforcing minority classes) is urgently needed.

🤔 Is this imbalance collection bias or reality?

The dominance of synthetic resin–vinyl (79,560 images) may genuinely reflect that vinyl waste is the most common in real industrial settings. If so, the imbalance faithfully mirrors reality, and forcibly balancing classes could produce a model disconnected from real-world distributions.

However, having a minimum class of only 20 images is less a reflection of reality than a collection gap. The ideal approach is to reference real-world distributions while ensuring every class has enough samples (at least several hundred) for the model to learn effectively.

Level 2 — DataLens Analysis (Wolfram ImageIdentify Net V2)

Analysis using a 1,280-dimensional general-purpose neural network. We examine how 72 waste classes are distributed in the feature space. Despite L1's severe imbalance, L2 geometry and distribution received a Good rating.

Industrial waste L2 PCA distribution

▲ Level 2 PCA Distribution — 72 classes naturally grouped into 5 clusters

Industrial waste L2 density topography

▲ Level 2 Overall Density Topography — 5 major clusters, bell-shaped uniform distribution

💡 L2 Key Finding — 5 Visual Groups of Waste: The general-purpose AI recognizes 72 waste types as 5 visual groups. Expected group composition: ① Metal containers ② Textile/fiber ③ Glass/ceramic ④ Plastic/vinyl ⑤ Construction debris. Interestingly, even a general-purpose AI can distinguish waste by material characteristics to some degree.

Per-Class Density Plots (L2)

Metal-Others L2 density Density
Metal-Others representative Actual
Metal-Others
Metal-Lacquer L2 density Density
Metal-Lacquer representative Actual
Metal-Lacquer
Textile-Others L2 density Density
Textile-Others representative Actual
Textile-Others
Textile-Fabric L2 density Density
Textile-Fabric representative Actual
Textile-Fabric
Textile-Sleeping Bag L2 density Density
Textile-Sleeping Bag representative Actual
Textile-Sleeping Bag
Textile-Tent L2 density Density
Textile-Tent representative Actual
Textile-Tent

▲ Each card — Left: L2 density distribution chart / Right: representative sample (actual image)

Level 3 — Domain-Specialized Analysis (136 Dimensions)

A domain-specialized lens optimized to 136 dimensions is applied. L2's 5 clusters are compressed into 3 in L3, providing clearer waste group distinctions. Both L3 geometry and distribution are Good — one of the few positive findings of this dataset.

Industrial waste L3 PCA distribution

▲ Level 3 PCA Distribution — 3-cluster structure in domain-optimized 136 dimensions

Industrial waste L3 density topography

▲ Level 3 Density Topography — High-density concentration confirmed in certain classes (Glass & Ceramic-Ceramic)

3 Waste Groups Discovered by L3

1

Rigid Waste Group

Metal Waste, Glass & Ceramic, Waste Concrete. Common traits: hard, glossy or opaque textures. Ceramic is the core high-density cluster of this group.

2

Flexible Waste Group

Textile Waste, Vinyl/Film, soft Synthetic Resin. Flexible and crumplable forms. Highest shape variation, concentrating outliers.

3

Mixed / Boundary Group

Paper Waste and composite waste. Located at the boundary of Groups 1 and 2, with mixed material characteristics.

Per-Class Density Plots (L3)

Metal-Others L3 density Density
Metal-Others representative Actual
Metal-Others
Metal-Lacquer L3 density Density
Metal-Lacquer representative Actual
Metal-Lacquer
Textile-Fabric L3 density Density
Textile-Fabric representative Actual
Textile-Fabric
Textile-Tent L3 density Density
Textile-Tent representative Actual
Textile-Tent
Textile-Sleeping Bag L3 density Density
Textile-Sleeping Bag representative Actual
Textile-Sleeping Bag
Textile-Others L3 density Density
Textile-Others representative Actual
Textile-Others

▲ Each card — Left: L3 density distribution chart / Right: representative sample (actual image)

Outlier Analysis — Why Is Ceramic Typical and Plastic Abnormal?

📏 What is "density"? Here, density does not refer to physical mass but to data concentration in feature space. High-density classes have similar images that AI can easily classify — but excessively high density suggests duplicate images. Low-density classes have such diverse images that AI struggles to find consistent patterns — these are the "outliers."

🎯 High Density — The Most "Typical" Waste for AI (Glass & Ceramic-Ceramic)

Ceramic (도자기류) fragments rank overwhelmingly #1 across the entire dataset with a density of 2.13. With their consistent form (broken fragments) and distinctive glossy texture, ceramic is the waste AI can classify with the highest confidence.

Ceramic high-density sample 1
Ceramic (density 2.13) 🔥
Ceramic high-density sample 2
Ceramic (density 2.13)
Ceramic high-density sample 3
Ceramic (density 2.11)
Ceramic high-density sample 4
Ceramic (density 2.08)
Ceramic high-density sample 5
Ceramic (density 2.07)
Ceramic high-density sample 6
Ceramic (density 2.04)
💡 Why Ceramic Is Typical: Waste ceramic fragments repeat a consistent pattern of broken shapes against white/beige backgrounds. Images shot consecutively at the same locations (X307, X024...) on the same day form high-density clusters. This may not be true "typicality" but rather the result of duplicate shooting — which is why Data Diet is needed.

⚠️ Low Density — The Most Confusing Outliers for AI

Synthetic Resin-Plastic (합성수지-플라스틱) and Metal-Paint Can (금속류-페인트통) dominate the low-density rankings. These are waste types with extremely diverse shapes or easily confused visually with other classes.

Synthetic Resin-Plastic low density
Plastic (density 0.36) 🔴
Metal-Paint Can low density
Metal Paint Can (density 0.36)
Metal-Others low density
Metal-Others (density 0.38)
Metal-Paint Can low density 2
Metal Paint Can (density 0.38)
Paper-Wallpaper low density
Paper-Wallpaper (density 0.39)
Ceramic low-density outlier
Ceramic outlier (density 0.39)

🔄 The Two Most Different Waste Types — Glass & Ceramic vs Window Glass / Waste Concrete

The farthest pair in L3: Glass & Ceramic-Others and Glass & Ceramic-Window Glass. Despite belonging to the same "Glass & Ceramic" major category, their detailed forms are extremely different.

Glass & Ceramic-Others
Glass & Ceramic-Others

Irregular glass fragments, transparent/translucent

Glass & Ceramic-Window Glass
Glass & Ceramic-Window Glass

Large flat glass panels, regular shape

Waste Concrete-Sewer Pipe
Waste Concrete-Sewer Pipe

Cylindrical structure, gray texture

Glass & Ceramic-Window Glass 2
Glass & Ceramic-Window Glass

Transparent large glass panel, can be confused with structures

💡 What this means for data quality:

The existence of an extremely different pair within the same "Glass/Ceramics" major category is a positive signal that the classification taxonomy is appropriately granular. However, from an AI model perspective, training on major categories alone will inevitably cause confusion — fine-grained class labels are essential. These extreme pairs also flag boundary classes that require special attention during data augmentation — improper augmentation can actually degrade model performance.

Improvement Recommendations — From 51 to 70+

🥗

① Data Diet

Remove similar images shot consecutively at the same location on the same day. As seen in the ceramic high-density cluster, images with density above 2.0 are likely duplicates.

Expected Effect: L1 Statistics 'Poor' → 'Fair'

💪

② Data Bulk-up

Add images from the minimum class (20 images) through mid-range classes. Goal: Bring the minimum class up to at least 50% of the average (approx. 5,000 images). Both data augmentation and additional photography can be used.

Expected Effect: L1 Class Balance 'Poor' → 'Fair~Good'

🎯 Practical Priority Suggestions

  1. Remove duplicate images (Diet) — Quick quality improvement
  2. Focus on reinforcing classes with 20–100 images (Bulk-up)
  3. Expand diversity of low-density outlier classes (Plastic, Metal-Paint Can)
  4. Standardize resolution — Unify camera equipment and shooting distance guidelines

Conclusion — The Potential and Challenges of 1 Million Images

The Industrial Waste Image dataset is impressive in scale. Its comprehensive coverage of 1 million images across 72 classes demonstrates the Korean government's commitment to AI data investment. However, the DataClinic score of 51 reveals a sobering reality — scale does not guarantee quality.

The 3,978x class imbalance and excess of similar images risk creating a biased AI if used directly for model training. But the good L2/L3 distribution structure means the data structure itself is sound — improvement through Diet and Bulk-up is entirely achievable.

Waste classification AI is a core technology for carbon neutrality and the circular economy. If this dataset is refined one step further, it can become a powerful foundation for automated waste classification and recycling AI at Korean industrial sites.

Industrial Waste Image Key Summary

51
DataClinic Overall
72
Waste Classes
1M
Image Scale
3,978x
Class Imbalance

Original DataClinic Report: dataclinic.ai/en/report/131 · AI Hub Original Data: AIHub #137 · Commercial use allowed

🔗 Related Datasets — Complementary Waste & Environmental Data

Household Waste Images (AI Hub)

Household and commercial waste images. Comparing with industrial waste reveals differences in classification schemes and image characteristics.

Search on AI Hub →

TrashNet (Stanford)

2,527 images across 6 recyclable categories. Small in scale but well-balanced — a stark contrast to this dataset's imbalance issues.

GitHub →

TACO (Trash Annotations in Context)

Real-environment litter images with segmentation annotations. Ideal for training object detection models.

tacodataset.org →

WasteNet (Kaggle)

Recyclable vs. non-recyclable binary classification dataset. Useful for quick baseline models and transfer learning experiments.

Search on Kaggle →