Jasonschwartzman/children-drawings-classifier-daniel-fraidin

🎨 Children's Drawings AI Classification System

Advanced AI system for analyzing children's drawings to identify potential psychological indicators using multiple state-of-the-art deep learning models.

🎯 Features

• —Binary Classification: Normal vs Problematic detection (up to 87% accuracy)
• —Multi-label Classification: Identifies 4 specific psychological indicators
• —Multiple Models: VGG19 and ConvNeXtV2 architectures
• —Comprehensive Analysis: Ensemble predictions from multiple models
• —Professional Reporting: Detailed analysis with recommendations

🤖 Models

Binary Classification Models

• —VGG19 Binary: 87.18% accuracy - Best for general screening
• —ConvNeXtV2 Binary: 85% accuracy - Lightweight and efficient

Multi-label Classification Models

• —VGG19 Multi-label: Traditional approach for indicator analysis
• —ConvNeXtV2 Multi-label: 67% avg F1-score - Superior detailed analysis

🏷️ Psychological Indicators Analyzed

1. 1.Problematic limbs: Issues with arm/leg representation and positioning
2. 2.Proportions: Disproportionate body parts indicating developmental concerns
3. 3.Missing body parts: Omitted essential body elements (head, limbs, torso)
4. 4.Distorted human figure: Overall figure distortion and unusual representations

📊 Usage

1. 1.Upload Image: Select a children's drawing (PNG, JPG, JPEG)
2. 2.Choose Analysis: Complete analysis, Binary only, or Multi-label only
3. 3.Select Model: ConvNeXtV2 (Modern) or VGG19 (Proven)
4. 4.Get Results: Comprehensive AI analysis with professional insights

⚠️ Important Disclaimers

• —Research Use Only: This tool is designed for research and educational purposes
• —Not Diagnostic: Should never be used as sole basis for psychological diagnosis
• —Professional Consultation: Always involve qualified mental health professionals
• —Ethical Use: Ensure proper consent and ethical guidelines when analyzing children's drawings

🔬 Technical Details

• —Training Data: 908 binary samples, 461 multilabel samples (augmented to 1816+)
• —Validation: Rigorous train/validation/test splits with stratification
• —Data Augmentation: Comprehensive augmentation (rotation, zoom, brightness, flipping)
• —Performance: Cross-validated training with early stopping and model checkpointing

📖 Citation

If you use this system in research, please cite appropriately and acknowledge the AI-assisted analysis nature.

Developed with advanced deep learning techniques for psychological research applications