smartdownsample 1.7.2

Smart image downsampling for image classification datasets

smartdownsample

Intelligent image downsampling with enhanced visual similarity detection

SmartDownsample uses multi-dimensional visual features to sample maximally diverse images from large collections. Optimized for camera trap data with center-focused animal detection, color/brightness separation, and natural folder ordering.

Installation

pip install smartdownsample

Features

• 🎯 Multi-dimensional visual similarity - Separates by structure, color, and brightness
• 🐾 Camera trap optimized - Center-focused detection for animals
• 🎨 Color-aware bucketing - Separates grayscale from color images
• 💡 Brightness distinction - Groups dark vs bright scenes
• 📊 Manageable grouping - Creates 16-32 meaningful buckets (not hundreds)
• ⚡ Still fast - Enhanced features with minimal speed impact
• 🎲 Reproducible - Set seed for consistent results
• 📈 Built-in visualization - Thumbnails grids and distribution charts

Usage

from smartdownsample import sample_diverse

# Basic usage - intelligent visual diversity
selected = sample_diverse(
    image_paths=my_image_list,
    target_count=1000
)

# Full feature usage with visualization
selected = sample_diverse(
    image_paths=my_camera_trap_images,
    target_count=1000,
    hash_size=8,                # Perceptual hash size (8 recommended)
    n_workers=4,                # Parallel workers
    show_progress=True,         # Progress bars
    random_seed=42,             # Reproducible results
    show_summary=True,          # Text statistics  
    show_distribution=True,     # Bucket distribution chart
    show_thumbnails=True        # 10x10 thumbnail grids per bucket
)

print(f"Selected {len(selected)} images from {len(buckets)} visual similarity groups")

Visualization Options

The algorithm includes three built-in visualization modes to understand bucket quality:

# 1. Text summary (show_summary=True) - Default
selected = sample_diverse(paths, target_count=1000, show_summary=True)
# Prints: bucket sizes, distribution stats, diversity metrics

# 2. Distribution chart (show_distribution=True) 
selected = sample_diverse(paths, target_count=1000, show_distribution=True)  
# Shows: vertical bar chart of kept vs excluded per bucket

# 3. Thumbnail grids (show_thumbnails=True)
selected = sample_diverse(paths, target_count=1000, show_thumbnails=True)
# Shows: 10x10 grids of first 100 images from each bucket in square layout

# All visualizations together
selected = sample_diverse(paths, target_count=1000, 
                         show_summary=True, 
                         show_distribution=True, 
                         show_thumbnails=True)

How It Works

Multi-dimensional visual similarity algorithm optimized for camera trap data:

1. Multi-Feature Extraction

Each image is analyzed using 4 complementary visual features:

# For each image, compute:
1. DHash (8x8) → Structural patterns, edges, shapes
2. AHash (4x4) → Brightness distribution, contrast  
3. Color Variance → Separates grayscale from colorful images
4. Overall Brightness → Separates dark from bright scenes

2. Center-Focused Animal Detection

For camera trap data where animals are typically centered:

# From 8x8 DHash (64 bits), strategically sample center positions:
center_indices = [27, 36]  # Center-left and center-right positions
# Bit 27: Detects vertical edges (animal body/legs)  
# Bit 36: Detects horizontal edges (animal head/back)

3. Smart Bucket Key Creation

Combine features into meaningful visual groups (max 32 buckets):

bucket_key = (
    structure_bit_27,     # Center-left animal features (0 or 1)
    structure_bit_36,     # Center-right animal features (0 or 1)  
    brightness_pattern,   # AHash brightness pattern (0 or 1)
    color_type,          # Grayscale=0, Color=1
    brightness_level     # Dark=0, Bright=1
)
# Results in 2×2×2×2×2 = 32 maximum buckets

4. Diversity-Preserving Selection

# Phase 1: Ensure diversity - sample from every bucket
# Phase 2: Fill remaining quota proportionally from largest buckets
# Within buckets: Natural sort preserves camera/folder structure

Example output buckets for camera trap data:
• Bucket 1: Dark grayscale deer (vertical edges)
• Bucket 2: Bright color birds (horizontal patterns) 
• Bucket 3: Grayscale empty frames (low structure)
• Bucket 4: Color daytime mammals (mixed patterns)

Algorithm Benefits

Visual Similarity Improvements

• Better separation: Color vs grayscale images grouped separately
• Animal-focused: Center-positioned features detect different animal poses/species
• Brightness aware: Day vs night scenes properly distinguished
• Structure sensitive: Different animal orientations and camera angles detected
• Manageable buckets: 16-32 meaningful groups instead of random mixing

Performance Characteristics

• Still fast: Multi-feature extraction adds minimal overhead (~20% slower)
• Linear scaling: O(n) complexity maintained across all features
• Memory efficient: Features computed on-the-fly, not stored
• Parallel processing: Hash computation parallelized across workers
• Smart bucket counts: Never creates excessive micro-buckets

Camera Trap Optimizations

• Natural sorting: Preserves camera/folder structure (CAM01_IMG_001.jpg → CAM01_IMG_010.jpg)
• Center detection: Focus on image center where animals appear
• Scene variety: Separates empty frames, single animals, multiple animals
• Lighting diversity: Day/night scenes properly represented
• Color preservation: IR grayscale vs color daylight images distinguished

Comparison with Other Methods

Method	Bucket Quality	Speed	Animal Detection	Color Separation
Random sampling	None	Fastest	No	No
Single DHash	Poor mixing	Fast	No	No
smartdownsample v1.6+	Excellent	Fast+	Yes	Yes
Complex ML clustering	Perfect	Very Slow	Depends	Yes

Real Results: Camera Trap Dataset

Before (v1.5): 495 buckets, color/grayscale mixed randomly in each bucket
After (v1.6+): 32 buckets, clear separation:

• Bucket 1: Grayscale deer images (IR night camera)
• Bucket 2: Color bird images (daylight camera)
• Bucket 3: Dark empty frames (nighttime)
• Bucket 4: Bright color mammals (sunny daytime)

Performance: Only ~20% slower than single-hash method, dramatically better visual grouping.

Performance

Task	Time
100 from 1,000	<5 sec
900 from 1,000	<5 sec
1,000 from 24,000	~30 sec
23,000 from 24,000	~30 sec
Any ratio	Fast ✓

Parameters

Parameter	Default	Description
image_paths	Required	List of image file paths (str or Path objects)
target_count	Required	Exact number of images to select
hash_size	8	Perceptual hash size - 8 recommended for good speed/quality balance
n_workers	4	Number of parallel workers for hash computation
show_progress	True	Display progress bars during processing
random_seed	42	Random seed for reproducible bucket selection
show_summary	True	Print bucket statistics and distribution summary
show_distribution	False	Show bucket distribution bar chart (requires matplotlib)
show_thumbnails	False	Show 10x10 thumbnail grids for each bucket (requires matplotlib)

Parameter Recommendations

For camera trap data:

• hash_size=8: Optimal balance of speed and animal detection quality
• show_thumbnails=True: Essential for validating bucket quality
• show_summary=True: Understand bucket size distribution

For other image types:

• hash_size=6: Faster processing, may reduce center-detection accuracy
• hash_size=10: Slower but more detailed structural analysis

Technical Details

Hash Features Explained

DHash (Difference Hash)

• Detects structural patterns, edges, object boundaries
• 8×8 hash = 64 bits representing horizontal gradients
• Center bits (positions 27, 36) focus on animal detection
• Fast computation: resize → grayscale → compare adjacent pixels

AHash (Average Hash)

• Detects brightness patterns and contrast distribution
• 4×4 hash = 16 bits representing above/below average brightness
• Used for distinguishing lighting conditions
• Complements DHash with tonal information

Color Variance

• Separates grayscale (IR cameras) from color (daylight cameras)
• Computed as variance of RGB channel means
• Threshold: variance < 100 = grayscale, ≥ 100 = color

Overall Brightness

• Separates dark (nighttime) from bright (daytime) scenes
• Computed as mean pixel value across all channels
• Threshold: brightness < 128 = dark, ≥ 128 = bright

Performance Notes

• Multi-feature extraction adds ~20% processing time vs single hash
• Parallel hash computation scales linearly with worker count (up to CPU cores)
• Memory usage remains O(1) - features computed on-demand
• Bucket creation is O(n) - no expensive similarity comparisons

License

MIT License – see LICENSE file.