# Constrained Fuzzy OCR - The Three-Tier OCR Pipeline
2026-01-07T18:00
[Part 4: Image Intelligence](/blog/constrained-fuzzy-image-intelligence) introduced the ImageSummarizer wave architecture and the broader patterns. This article deep-dives into the **OCR subsystem**—three tiers of text extraction, intelligent routing, and the filmstrip optimization that achieves 30× token reduction for animated GIFs.
**Why a separate article?** The OCR pipeline evolved from "Tesseract with Vision LLM fallback" to a sophisticated three-tier system with ML-based OCR, multi-frame voting, text-only strip extraction, and cost-aware routing. It's complex enough to warrant its own detailed breakdown.
**Related articles**:
- [Part 1: Constrained Fuzziness Pattern](/blog/constrained-fuzziness-pattern) - The foundational pattern
- [Part 2: Constrained Fuzzy MoM](/blog/constrained-mom-mixture-of-models) - Multiple model coordination
- [Part 3: Context Dragging](/blog/constrained-fuzzy-context-dragging) - Memory and time
- [Part 4: Image Intelligence](/blog/constrained-fuzzy-image-intelligence) - Full wave architecture overview
- [DocSummarizer](/blog/building-a-document-summarizer-with-rag) - Document analysis with similar patterns
- [DataSummarizer](/blog/datasummarizer-how-it-works) - Data profiling with determinism-first approach
[TOC]
---
## The Problem: Text Extraction is Hard
OCR on real-world images fails in predictable ways:
- **Stylized fonts**: Tesseract trained on standard fonts, fails on decorative text
- **Noisy GIFs**: Frame compression artifacts, jitter, subtitle changes
- **Low contrast**: Dark text on dark backgrounds
- **Rotated text**: Non-horizontal text angles
- **Mixed content**: Screenshots with multiple text regions
- **API costs**: Vision LLM calls are expensive ($0.001-0.01 per image)
Traditional approach: "Run Tesseract, if it fails use Vision LLM"
**Problem**: This either misses stylized text (Tesseract fails) or costs too much (always use Vision LLM).
**Solution**: Add a middle tier (Florence-2 ONNX) that handles stylized fonts locally, escalating to Vision LLM only when both local methods fail.
---
## The Three-Tier OCR Architecture
The system runs waves in priority order (higher number = later execution):
```
Wave Priority Order:
40: TextLikelinessWave → Heuristic text detection
50: OcrWave → Tesseract OCR (if text-likely)
51: MlOcrWave → Florence-2 ML OCR (if Tesseract low confidence)
55: Florence2Wave → Florence-2 captions (optional)
80: VisionLlmWave → Vision LLM (escalation)
```
### Tier 1: Tesseract (Traditional OCR)
| Priority | Speed | Cost | Best For | Limitations |
|----------|-------|-------|----------|------------|
| **50** | ~50ms | Free | Clean text, high contrast, standard fonts | Stylized fonts, low quality, rotated text |
**Signals emitted:**
- `ocr.text` - Extracted text
- `ocr.confidence` - Tesseract mean confidence score
### Tier 2: Florence-2 ONNX (ML OCR)
| Priority | Speed | Cost | Best For | Limitations |
|----------|-------|-------|----------|------------|
| **51** | ~200ms | Free | Stylized fonts, memes, decorative text | Complex charts, rotated text |
**Signals emitted:**
- `ocr.ml.text` - Single-frame Florence-2 OCR
- `ocr.ml.multiframe_text` - Multi-frame GIF text (preferred for animations)
- `ocr.ml.confidence` - Model confidence score
### Tier 3: Vision LLM (Cloud Fallback)
| Priority | Speed | Cost | Best For | Constraints |
|----------|-------|-------|----------|------------|
| **80** | ~1-5s | $0.001-0.01 | Everything, especially complex scenes | Must respect deterministic signals |
**Signals emitted:**
- `ocr.vision.text` - Vision LLM OCR text extraction
- `ocr.vision.confidence` - LLM confidence (typically 0.95)
- `caption.text` - Optional descriptive caption (separate from OCR)
---
## The ONNX Arsenal: Local ML Models
Before diving into the three OCR tiers, let's cover the **deterministic ML models** that power the system. All models run locally via ONNX Runtime—no API calls, no cloud dependencies, no costs.
### Why ONNX?
- **Runs locally**: No API keys, no network latency, no recurring costs
- **Deterministic**: Same input = same output (no sampling/temperature randomness)*
- **Fast**: Hardware-accelerated (CPU/GPU), optimized inference
- **Portable**: Works on Windows, Linux, macOS
- **Auto-downloaded**: First run downloads models automatically
\* *Minor caveat: GPU execution providers can introduce negligible floating-point nondeterminism. The signal contract (confidence thresholds, routing logic) remains fully deterministic.*
### The Five ONNX Models
> **Note**: Sizes are approximate and vary by variant/quantization. Typical download sizes shown below.
| Model | Approx. Size | Purpose | Speed | Model Type |
|-------|--------------|---------|-------|------------|
| **EAST** | ~100MB | Scene text detection | ~20ms | Text detection |
| **CRAFT** | ~150MB | Character-region text detection | ~30ms | Text detection |
| **Florence-2** | ~250MB | OCR + captioning | ~200ms | Vision-language |
| **Real-ESRGAN** | ~60MB | 4× super-resolution upscaling | ~500ms | Image enhancement |
| **CLIP** | ~600MB | Semantic embeddings | ~100ms | Multimodal embedding |
**Total disk space**: ~1.0-1.5GB depending on model variants chosen.
---
### 1. EAST: Scene Text Detection
**Efficient and Accurate Scene Text Detector** - finds text regions in natural scenes.
```csharp
// EAST detects text bounding boxes with confidence scores
var result = await textDetector.RunEastDetectionAsync(imagePath);
// Output: List of BoundingBox with coordinates + confidence
// Example: [BoundingBox(x1:50, y1:100, x2:300, y2:150, confidence:0.92)]
```
**How it works**:
- Deep learning model trained on scene text datasets
- Outputs score map (confidence) + geometry map (box coordinates)
- Handles rotated text, multi-scale text
- Uses Non-Maximum Suppression (NMS) to merge overlapping boxes
**Why deterministic?**
- No randomness in inference (frozen weights)
- Same image → same bounding boxes
- Confidence scores are reproducible
- Escalation thresholds are config-driven (e.g., `< 0.5 → escalate`)
**Technical details**:
```csharp
// EAST preprocessing (from implementation)
- Input size: 320×320 (must be multiple of 32)
- Format: BGR with mean subtraction [123.68, 116.78, 103.94]
- Output stride: 4 (downsampled 4×)
- Score threshold: 0.5
- NMS IoU threshold: 0.4
```
**Example output**:
```
Input: meme.png (800×600)
EAST detection: 15 text regions found
Region 1: (50, 480, 750, 580) - confidence 0.87 [bottom subtitle area]
Region 2: (100, 50, 300, 90) - confidence 0.62 [top text]
Region 3: ...
Route decision: ANIMATED (subtitle pattern in bottom 30%)
```
---
### 2. CRAFT: Character Region Awareness
**Character-level text detection** - excels at curved, artistic, and stylized text.
```csharp
// CRAFT finds character-level regions, then groups into words
var result = await textDetector.RunCraftDetectionAsync(imagePath);
// Better than EAST for: decorative fonts, curved text, logos
```
**How it works**:
- Detects individual character regions (more granular than EAST)
- Uses affinity score to group characters into words
- Flood-fill algorithm finds connected text components
- Handles curved text that EAST misses
**When CRAFT is used**:
1. EAST is unavailable or failed
2. Image has artistic/decorative fonts (auto-detected)
3. User explicitly selects CRAFT detector
**Technical details**:
```csharp
// CRAFT preprocessing
- Max dimension: 1280px (maintains aspect ratio)
- Format: RGB normalized with ImageNet stats
- Mean: [0.485, 0.456, 0.406]
- Std: [0.229, 0.224, 0.225]
- Output stride: 2 (downsampled 2×)
- Threshold: 0.4 for character regions
```
**EAST vs CRAFT comparison**:
| Feature | EAST | CRAFT |
|---------|------|-------|
| Detection level | Word/line | Character |
| Speed | ~20ms | ~30ms |
| Best for | Standard text, subtitles | Decorative fonts, logos |
| Curved text | Limited | Excellent |
| Model size | 100MB | 150MB |
---
### 3. Real-ESRGAN: Super-Resolution Upscaling
**Enhances low-quality images before OCR** - 4× upscaling for blurry/small text.
```csharp
// Upscale low-quality image before running OCR
if (quality.Sharpness < 30) // Laplacian variance threshold
{
var upscaled = await esrganService.UpscaleAsync(imagePath, scale: 4);
// Now run OCR on the enhanced image
}
```
**When it's used**:
- Image sharpness < 30 (Laplacian variance)
- Text regions detected but very small (< 20px height)
- OCR confidence low but text regions present
- User explicitly requests upscaling
**Example**:
```
Input: 100×75 screenshot with tiny text
Laplacian variance: 18 (very blurry)
ESRGAN: Upscale to 400×300 (~500ms)
New Laplacian variance: 87 (sharp)
OCR: Tesseract confidence: 0.92 (vs 0.42 before upscaling)
Text: "Click here to continue" (vs garbled before)
```
**Technical details**:
```csharp
// Real-ESRGAN processing
- Input: Any size (processed in 128×128 tiles if large)
- Output: 4× scaled (200×150 → 800×600)
- Model: x4plus variant (general photos)
- Processing: ~500ms for 800×600 image
- Memory: ~2GB peak (tiles reduce this)
```
**Token economics**:
```
Scenario: Screenshot with tiny text
Option 1: Send low-res to Vision LLM
Image: 100×75 = ~20 tokens
LLM can't read tiny text → fails
Cost: $0.0002 (wasted)
Option 2: Upscale with ESRGAN, use Tesseract
ESRGAN: Free (local), 500ms
Tesseract: Free (local), 50ms
Success: 92% confidence
Cost: $0
Result: ESRGAN + local OCR beats Vision LLM for low-res images
```
---
### 4. CLIP: Semantic Embeddings
**Multimodal embeddings for semantic image search** - projects images and text into shared vector space.
```csharp
// Generate embedding for semantic search
var embedding = await clipService.GenerateEmbeddingAsync(imagePath);
// Returns: float[512] vector
// Later: semantic search across thousands of images
var similarImages = await vectorDb.SearchAsync(queryEmbedding, topK: 10);
```
**How it works**:
- CLIP ViT-B/32 visual encoder (350MB)
- Projects images to 512-dimensional vectors
- Trained to align with text descriptions
- Enables "find images like this" without keywords
**Use cases**:
- Semantic image search in RAG systems
- Duplicate detection (even if edited/cropped)
- Content-based clustering
- Similar image recommendations
**Technical details**:
```csharp
// CLIP visual encoder
- Model: ViT-B/32 (Vision Transformer)
- Input: 224×224 RGB (center crop + resize)
- Output: 512-dimensional embedding
- Normalized: L2 norm = 1.0
- Speed: ~100ms per image
```
**Example**:
```
Input images:
cat_on_couch.jpg → [0.23, -0.51, 0.88, ...]
dog_on_couch.jpg → [0.19, -0.48, 0.91, ...]
car_photo.jpg → [-0.67, 0.33, -0.12, ...]
Query: "animals on furniture"
Text embedding → [0.21, -0.50, 0.89, ...]
Cosine similarity:
cat_on_couch: 0.94 (very similar!)
dog_on_couch: 0.91 (similar)
car_photo: 0.12 (not similar)
Result: Returns cat and dog images
```
---
### 5. Florence-2: Vision-Language Model (Covered in Tier 2)
See Tier 2 section for full details on Florence-2 ONNX OCR and captioning.
---
### Auto-Download System
All models are downloaded automatically on first use:
```bash
$ imagesummarizer image.png --pipeline auto
[First run]
Downloading EAST scene text detector (~100MB)...
Progress: ████████████████████ 100% (102.4 MB)
Downloading Florence-2 base model (~250MB)...
Progress: ████████████████████ 100% (248.7 MB)
Downloading CLIP ViT-B/32 visual (~350MB)...
Progress: ████████████████████ 100% (347.2 MB)
Models saved to: ~/.mostlylucid/models/
Total disk space: 1.16 GB
[Subsequent runs]
All models cached, analysis starts immediately
```
**Graceful degradation**:
```csharp
// If ONNX model download fails, system falls back gracefully
EAST unavailable → Try CRAFT → Fall back to Tesseract PSM
Real-ESRGAN unavailable → Skip upscaling, use original image
CLIP unavailable → Skip embeddings, OCR still works
Florence-2 unavailable → Use Tesseract → Vision LLM escalation
```
Every ONNX model failure is logged with fallback path, ensuring the system never crashes due to missing models.
---
### Why This Matters
> **Pricing note**: Cost examples below use illustrative pricing (~$0.005/image for Vision LLM). Actual API costs vary by provider and model. The core insight—local processing eliminates most API calls—holds regardless of specific pricing.
**Without ONNX models** (baseline):
```
Every image → Send to Vision LLM
Cost: ~$0.005/image (example pricing)
Time: ~2s network + inference
100 images = ~$0.50, ~200s
```
**With ONNX models** (local-first):
```
85 images → EAST + Florence-2 (local)
Cost: $0
Time: ~200ms
10 images → EAST + Tesseract (local)
Cost: $0
Time: ~50ms
5 images → EAST + Vision LLM (escalation)
Cost: ~$0.025 (5 × $0.005)
Time: ~2s each
100 images = ~$0.025, ~30s total
```
**Savings**: ~95% cost reduction, ~85% faster, **deterministic routing**.
The ONNX models transform the system from "probabilistic all the way down" to "deterministic foundation + probabilistic escalation only when needed."
---
## Tier 1: Tesseract OCR
The baseline. Fast, deterministic, works great for clean text.
```csharp
public class OcrWave : IAnalysisWave
{
public string Name => "OcrWave";
public int Priority => 60; // After color/identity
public async Task> AnalyzeAsync(
string imagePath,
AnalysisContext context,
CancellationToken ct)
{
var signals = new List();
// Get preprocessed image from cache
var image = context.GetCached>("image");
// Run Tesseract OCR
using var engine = new TesseractEngine(@"./tessdata", "eng", EngineMode.Default);
using var page = engine.Process(image);
var text = page.GetText();
var confidence = page.GetMeanConfidence();
signals.Add(new Signal
{
Key = "ocr.text", // Tesseract OCR result
Value = text,
Confidence = confidence,
Source = Name,
Tags = new List { "ocr", "text" },
Metadata = new Dictionary
{
["engine"] = "tesseract",
["mean_confidence"] = confidence,
["word_count"] = text.Split(' ').Length
}
});
signals.Add(new Signal
{
Key = "ocr.confidence",
Value = confidence,
Confidence = 1.0,
Source = Name
});
return signals;
}
}
```
**Key signals**:
- `ocr.full_text` - The extracted text
- `ocr.early_exit` - Signal to skip Tier 2/3 if confidence is high
- Confidence score drives escalation decisions
---
## Tier 2: Florence-2 ONNX
Microsoft's Florence-2 is a vision-language model that excels at dense captioning and OCR. The ONNX version runs locally with no API costs.
### Why Florence-2?
- **Better than Tesseract for stylized fonts**: Handles decorative text, memes, logos
- **Faster than Vision LLM**: ~200ms vs 1-5s
- **Free**: Runs locally, no API key required
- **Multimodal understanding**: Can extract text in context (e.g., speech bubbles)
### Implementation
```csharp
public class MlOcrWave : IAnalysisWave
{
private readonly Florence2OnnxModel _model;
public string Name => "MlOcrWave";
public int Priority => 51; // Runs AFTER Tesseract (priority 50)
public async Task> AnalyzeAsync(
string imagePath,
AnalysisContext context,
CancellationToken ct)
{
var signals = new List();
// Check if Tesseract already succeeded with high confidence
var tesseractConfidence = context.GetValue("ocr.confidence");
if (tesseractConfidence >= 0.95)
{
signals.Add(new Signal
{
Key = "ocr.ml.skipped", // Consistent namespace: ocr.ml.*
Value = true,
Confidence = 1.0,
Source = Name,
Metadata = new Dictionary
{
["reason"] = "tesseract_high_confidence",
["tesseract_confidence"] = tesseractConfidence
}
});
return signals;
}
// Run Florence-2 OCR
var result = await _model.ExtractTextAsync(imagePath, ct);
signals.Add(new Signal
{
Key = "ocr.ml.text", // Florence-2 ML OCR text
Value = result.Text,
Confidence = result.Confidence,
Source = Name,
Tags = new List { "ocr", "text", "ml" },
Metadata = new Dictionary
{
["model"] = "florence2-base",
["inference_time_ms"] = result.InferenceTime,
["token_count"] = result.TokenCount
}
});
// For animated GIFs, extract all unique frames
if (context.GetValue("identity.frame_count") > 1)
{
var frameResults = await ExtractMultiFrameTextAsync(
imagePath,
maxFrames: 10,
ct);
signals.Add(new Signal
{
Key = "ocr.ml.multiframe_text",
Value = frameResults.CombinedText,
Confidence = frameResults.AverageConfidence,
Source = Name,
Metadata = new Dictionary
{
["frames_processed"] = frameResults.FrameCount,
["unique_text_segments"] = frameResults.UniqueSegments,
["deduplication_method"] = "levenshtein_85"
}
});
}
return signals;
}
}
```
### Multi-Frame GIF Processing
For animated GIFs, Florence-2 processes up to 10 sampled frames in parallel:
```csharp
private async Task ExtractMultiFrameTextAsync(
string imagePath,
int maxFrames,
CancellationToken ct)
{
// Load GIF and extract frames
using var image = await Image.LoadAsync(imagePath, ct);
var frames = new List>();
int frameCount = image.Frames.Count;
int step = Math.Max(1, frameCount / maxFrames);
for (int i = 0; i < frameCount; i += step)
{
frames.Add(image.Frames.CloneFrame(i));
}
// Process all frames in parallel (bounded concurrency to avoid thrashing)
var semaphore = new SemaphoreSlim(4); // Max 4 concurrent inferences
var tasks = frames.Select(async frame =>
{
await semaphore.WaitAsync(ct);
try
{
var result = await _model.ExtractTextAsync(frame, ct);
return result;
}
finally
{
semaphore.Release();
}
});
var results = await Task.WhenAll(tasks);
semaphore.Dispose();
// Deduplicate using Levenshtein distance
var uniqueTexts = DeduplicateByLevenshtein(
results.Select(r => r.Text).ToList(),
threshold: 0.85);
return new MultiFrameResult
{
CombinedText = string.Join("\n", uniqueTexts),
FrameCount = frames.Count,
UniqueSegments = uniqueTexts.Count,
AverageConfidence = results.Average(r => r.Confidence)
};
}
private List DeduplicateByLevenshtein(
List texts,
double threshold)
{
var unique = new List();
foreach (var text in texts)
{
bool isDuplicate = false;
foreach (var existing in unique)
{
var distance = LevenshteinDistance(text, existing);
var maxLen = Math.Max(text.Length, existing.Length);
var similarity = 1.0 - (distance / (double)maxLen);
if (similarity >= threshold)
{
isDuplicate = true;
break;
}
}
if (!isDuplicate)
{
unique.Add(text);
}
}
return unique;
}
```
**Example**: 93-frame GIF → 10 sampled frames → 2 unique text results
```
Frame 1-45: "I'm not even mad."
Frame 46-93: "That's amazing."
```
---
## The Routing Decision
OpenCV text detection (~5-20ms) determines which path to take:
```csharp
public class TextDetectionService
{
public TextDetectionResult DetectText(Image image)
{
// Use OpenCV EAST text detector
var (regions, confidence) = RunEastDetector(image);
return new TextDetectionResult
{
HasText = regions.Count > 0,
RegionCount = regions.Count,
Confidence = confidence,
Route = SelectRoute(regions, confidence, image)
};
}
private ProcessingRoute SelectRoute(
List regions,
double confidence,
Image image)
{
// No text detected
if (regions.Count == 0)
return ProcessingRoute.NoOcr;
// Animated GIF with subtitle pattern
if (image.Frames.Count > 1 && HasSubtitlePattern(regions))
return ProcessingRoute.AnimatedFilmstrip;
// High confidence, standard text
if (confidence >= 0.8 && HasStandardTextCharacteristics(regions))
return ProcessingRoute.Fast; // Florence-2 only
// Moderate confidence
if (confidence >= 0.5)
return ProcessingRoute.Balanced; // Florence-2 + Tesseract voting
// Low confidence, complex image
return ProcessingRoute.Quality; // Full pipeline + Vision LLM
}
private bool HasSubtitlePattern(List regions)
{
// Subtitles are typically in bottom 30% of frame
var bottomRegions = regions.Where(r =>
r.BoundingBox.Y > r.ImageHeight * 0.7);
return bottomRegions.Count() >= regions.Count * 0.5;
}
}
```
### Route Performance
| Route | Triggers When | Processing | Time | Cost |
|-------|---------------|------------|------|------|
| **FAST** | High confidence (>0.8), standard text | Florence-2 only | ~100ms | Free |
| **BALANCED** | Moderate confidence (0.5-0.8) | Florence-2 + Tesseract voting | ~300ms | Free |
| **QUALITY** | Low confidence (<0.5), complex | Multi-frame + Vision LLM | ~1-5s | $0.001-0.01 |
| **ANIMATED** | GIF with subtitle pattern | Text-only filmstrip | ~2-3s | $0.002-0.005 |
---
## Text-Only Strip Extraction
The breakthrough optimization for GIF subtitles: extract **only the text regions**, not full frames.
### The Problem
Traditional approach for a 93-frame GIF with subtitles:
```
Option 1: Process every frame
93 frames × 300×185 × ~150 tokens/frame = 13,950 tokens
Cost: ~$0.14 @ $0.01/1K tokens
Time: ~27 seconds
Option 2: Sample 10 frames
10 frames × 300×185 × ~150 tokens/frame = 1,500 tokens
Cost: ~$0.015
Time: ~3 seconds
Problem: Might miss subtitle changes
```
### The Solution: Text-Only Strips
Extract only the text bounding boxes, eliminating background pixels:
```
2 text regions × 250×50 × ~25 tokens/region = 50 tokens
Cost: ~$0.0005
Time: ~2 seconds
Token reduction: 30×
```
### Implementation
```csharp
public class FilmstripService
{
public async Task CreateTextOnlyStripAsync(
string imagePath,
CancellationToken ct)
{
using var gif = await Image.LoadAsync(imagePath, ct);
// 1. Detect subtitle region (bottom 30% of frames)
var subtitleRegion = DetectSubtitleRegion(gif);
// 2. Extract frames with text changes
var uniqueFrames = ExtractUniqueTextFrames(gif, subtitleRegion);
// 3. Extract tight bounding boxes around text
var textRegions = ExtractTextBoundingBoxes(uniqueFrames);
// 4. Create horizontal strip of text-only regions
var strip = CreateHorizontalStrip(textRegions);
return new TextOnlyStrip
{
Image = strip,
RegionCount = textRegions.Count,
TotalTokens = EstimateTokens(strip),
OriginalTokens = EstimateTokens(gif),
Reduction = CalculateReduction(strip, gif)
};
}
private Rectangle DetectSubtitleRegion(Image gif)
{
// Analyze bottom 30% of frame for text patterns
int subtitleHeight = (int)(gif.Height * 0.3);
int subtitleY = gif.Height - subtitleHeight;
return new Rectangle(0, subtitleY, gif.Width, subtitleHeight);
}
private List> ExtractUniqueTextFrames(
Image gif,
Rectangle subtitleRegion)
{
var uniqueFrames = new List>();
Image? previousFrame = null;
for (int i = 0; i < gif.Frames.Count; i++)
{
var frame = gif.Frames.CloneFrame(i);
var subtitleCrop = frame.Clone(ctx =>
ctx.Crop(subtitleRegion));
// Compare with previous frame
if (previousFrame == null ||
HasTextChanged(subtitleCrop, previousFrame, threshold: 0.05))
{
uniqueFrames.Add(subtitleCrop);
previousFrame = subtitleCrop;
}
}
return uniqueFrames;
}
private bool HasTextChanged(
Image current,
Image previous,
double threshold)
{
// Threshold bright pixels (white/yellow text on dark background)
var currentBright = CountBrightPixels(current);
var previousBright = CountBrightPixels(previous);
// Calculate Jaccard similarity of bright pixels
var intersection = currentBright.Intersect(previousBright).Count();
var union = currentBright.Union(previousBright).Count();
var similarity = union > 0 ? intersection / (double)union : 1.0;
// Text changed if similarity drops below threshold
return similarity < (1.0 - threshold);
}
// Helper type for bounding box + crop
private record TextCrop
{
public required Image CroppedImage { get; init; }
public required Rectangle Bounds { get; init; }
}
private List ExtractTextBoundingBoxes(
List> frames)
{
var textCrops = new List();
foreach (var frame in frames)
{
// Threshold to get text mask
var mask = ThresholdBrightPixels(frame, minValue: 200);
// Find connected components (text regions)
var components = FindConnectedComponents(mask);
// Get tight bounding box around all components
var bbox = GetTightBoundingBox(components);
// Add padding
bbox.Inflate(5, 5);
// Clone the region (dispose properly in production!)
var cropped = frame.Clone(ctx => ctx.Crop(bbox));
textCrops.Add(new TextCrop
{
CroppedImage = cropped,
Bounds = bbox
});
}
return textCrops;
}
private Image CreateHorizontalStrip(
List textCrops)
{
// Calculate strip dimensions
int totalWidth = textCrops.Sum(c => c.Bounds.Width);
int maxHeight = textCrops.Max(c => c.Bounds.Height);
// Create blank canvas
var strip = new Image(totalWidth, maxHeight);
// Paste text regions horizontally
int xOffset = 0;
foreach (var crop in textCrops)
{
strip.Mutate(ctx => ctx.DrawImage(
crop.CroppedImage,
new Point(xOffset, 0),
opacity: 1.0f));
xOffset += crop.Bounds.Width;
// Dispose crop after use (important!)
crop.CroppedImage.Dispose();
}
return strip;
}
}
```
### Visual Example
**Input**: anchorman-not-even-mad.gif (93 frames, 300×185)
**Processing**:
```
1. Detect subtitle region: bottom 30% (300×55)
2. Extract unique frames: 93 frames → 2 text changes
3. Extract tight bounding boxes:
- Frame 1-45: "I'm not even mad." → 252×49 bbox
- Frame 46-93: "That's amazing." → 198×49 bbox
4. Create horizontal strip: 450×49 total
```
**Output**: Text-only strip (450×49)
**Token Economics**:
- Full frames (10 sampled): 300×185 × 10 = ~1500 tokens
- OCR strip (2 frames): 300×185 × 2 = ~300 tokens
- **Text-only strip**: 450×49 = ~50 tokens
**30× reduction** while preserving all subtitle text.
---
## Tier 3: Vision LLM Escalation
When both Tesseract and Florence-2 fail or produce low-confidence results, escalate to a Vision LLM (GPT-4o, Claude 3.5 Sonnet, Gemini Pro Vision, or Ollama models like minicpm-v).
### The Quality Gate
```csharp
public class OcrQualityWave : IAnalysisWave
{
private readonly SpellChecker _spellChecker;
public string Name => "OcrQualityWave";
public int Priority => 58; // After Florence-2 and Tesseract
public async Task> AnalyzeAsync(
string imagePath,
AnalysisContext context,
CancellationToken ct)
{
var signals = new List();
// Get best OCR result from earlier waves (priority order)
string? ocrText =
context.GetValue("ocr.ml.text") ?? // Florence-2 (priority 51)
context.GetValue("ocr.text"); // Tesseract (priority 50)
if (string.IsNullOrWhiteSpace(ocrText))
{
signals.Add(new Signal
{
Key = "ocr.quality.no_text",
Value = true,
Confidence = 1.0,
Source = Name
});
return signals;
}
// Run spell check (deterministic quality assessment)
var spellResult = _spellChecker.CheckTextQuality(ocrText);
// Additional quality signals to avoid false positives
var alphanumRatio = CalculateAlphanumericRatio(ocrText); // Letters/digits vs junk
var avgTokenLength = CalculateAverageTokenLength(ocrText);
signals.Add(new Signal
{
Key = "ocr.quality.spell_check_score",
Value = spellResult.CorrectWordsRatio,
Confidence = 1.0,
Source = Name,
Metadata = new Dictionary
{
["total_words"] = spellResult.TotalWords,
["correct_words"] = spellResult.CorrectWords,
["garbled_words"] = spellResult.GarbledWords,
["alphanum_ratio"] = alphanumRatio,
["avg_token_length"] = avgTokenLength
}
});
// Deterministic escalation threshold
// NOTE: Spellcheck alone can false-trigger on proper nouns, memes, brand names.
// Use additional signals (alphanum ratio, token length) to reduce false escalations.
bool isGarbled = spellResult.CorrectWordsRatio < 0.5 &&
alphanumRatio > 0.7; // Mostly valid characters, just not in dictionary
signals.Add(new Signal
{
Key = "ocr.quality.is_garbled",
Value = isGarbled,
Confidence = 1.0,
Source = Name
});
// Signal Vision LLM escalation
if (isGarbled)
{
signals.Add(new Signal
{
Key = "ocr.quality.escalation_required",
Value = true,
Confidence = 1.0,
Source = Name,
Tags = new List { "action_required", "escalation" },
Metadata = new Dictionary
{
["reason"] = "spell_check_below_threshold",
["quality_score"] = spellResult.CorrectWordsRatio,
["threshold"] = 0.5,
["target_tier"] = "vision_llm"
}
});
// Cache garbled text for Vision LLM to access
context.SetCached("ocr.garbled_text", ocrText);
}
return signals;
}
}
```
**Escalation is deterministic**: spell check score < 50% → escalate. No probabilistic judgment.
### Vision LLM with Filmstrip
When escalation is triggered for animated GIFs, use the text-only strip:
```csharp
public class VisionLlmWave : IAnalysisWave
{
private readonly IVisionLlmClient _client;
public string Name => "VisionLlmWave";
public int Priority => 50;
public async Task> AnalyzeAsync(
string imagePath,
AnalysisContext context,
CancellationToken ct)
{
var signals = new List();
// Check if escalation is required
var escalationRequired = context.GetValue(
"ocr.quality.escalation_required");
if (!escalationRequired)
{
signals.Add(new Signal
{
Key = "vision.llm.skipped",
Value = true,
Confidence = 1.0,
Source = Name,
Metadata = new Dictionary
{
["reason"] = "no_escalation_required"
}
});
return signals;
}
// For animated GIFs, use text-only strip
string imageToProcess = imagePath;
bool usedFilmstrip = false;
if (context.GetValue("identity.frame_count") > 1)
{
var filmstrip = await CreateTextOnlyStripAsync(imagePath, ct);
imageToProcess = filmstrip.Path;
usedFilmstrip = true;
signals.Add(new Signal
{
Key = "vision.filmstrip.created",
Value = true,
Confidence = 1.0,
Source = Name,
Metadata = new Dictionary
{
["mode"] = "text_only",
["region_count"] = filmstrip.RegionCount,
["token_reduction"] = filmstrip.Reduction,
["original_tokens"] = filmstrip.OriginalTokens,
["final_tokens"] = filmstrip.TotalTokens
}
});
}
// Build constrained prompt
var prompt = BuildConstrainedPrompt(context);
// Call Vision LLM
var result = await _client.ExtractTextAsync(
imageToProcess,
prompt,
ct);
// Emit OCR text signal (Vision LLM tier)
signals.Add(new Signal
{
Key = "ocr.vision.text", // Vision LLM OCR result
Value = result.Text,
Confidence = 0.95, // High but not 1.0 - still probabilistic
Source = Name,
Tags = new List { "ocr", "vision", "llm" },
Metadata = new Dictionary
{
["model"] = result.Model,
["used_filmstrip"] = usedFilmstrip,
["inference_time_ms"] = result.InferenceTime,
["token_count"] = result.TokenCount,
["cost_usd"] = result.Cost
}
});
// Optionally emit caption if requested (separate from OCR)
if (result.Caption != null)
{
signals.Add(new Signal
{
Key = "caption.text", // Descriptive caption, not OCR
Value = result.Caption,
Confidence = 0.90,
Source = Name,
Tags = new List { "caption", "description" }
});
}
return signals;
}
private string BuildConstrainedPrompt(AnalysisContext context)
{
var sb = new StringBuilder();
sb.AppendLine("Extract all text from this image.");
sb.AppendLine();
sb.AppendLine("CONSTRAINTS:");
sb.AppendLine("- Only extract text that is actually visible");
sb.AppendLine("- Preserve formatting and line breaks");
sb.AppendLine("- If no text is present, return empty string");
sb.AppendLine();
// Add context from earlier waves
var garbledText = context.GetCached("ocr.garbled_text");
if (!string.IsNullOrEmpty(garbledText))
{
sb.AppendLine("CONTEXT:");
sb.AppendLine("Traditional OCR detected garbled text:");
sb.AppendLine($" \"{garbledText}\"");
sb.AppendLine("Use this as a hint for stylized or unusual fonts.");
sb.AppendLine();
}
sb.AppendLine("Return only the extracted text, no commentary.");
return sb.ToString();
}
}
```
---
## The Priority Chain
When all tiers complete, the final text selection uses a strict priority order:
```csharp
public static string? GetFinalText(DynamicImageProfile profile)
{
// Priority chain (highest to lowest quality)
// NOTE: This selects ONE source, but the ledger exposes ALL sources
// with confidence scores for downstream inspection
// 1. Vision LLM OCR (best for complex/garbled text)
var visionText = profile.GetValue("ocr.vision.text");
if (!string.IsNullOrEmpty(visionText))
return visionText;
// 2. Florence-2 multi-frame GIF OCR (best for animations)
var florenceMultiText = profile.GetValue("ocr.ml.multiframe_text");
if (!string.IsNullOrEmpty(florenceMultiText))
return florenceMultiText;
// 3. Florence-2 single-frame ML OCR (good for stylized fonts)
var florenceText = profile.GetValue("ocr.ml.text");
if (!string.IsNullOrEmpty(florenceText))
return florenceText;
// 4. Tesseract OCR (reliable for clean standard text)
var tesseractText = profile.GetValue("ocr.text");
if (!string.IsNullOrEmpty(tesseractText))
return tesseractText;
// 5. Fallback (empty)
return string.Empty;
}
```
**Each tier has known characteristics**:
| Source | Signal Key | Best For | Confidence | Cost | Speed |
|--------|------------|----------|------------|------|-------|
| Vision LLM OCR | `ocr.vision.text` | Complex charts, rotated text, garbled | 0.95 | $0.001-0.01 | ~1-5s |
| Florence-2 (GIF) | `ocr.ml.multiframe_text` | Animated GIFs with subtitles | 0.85-0.92 | Free | ~200ms |
| Florence-2 (single) | `ocr.ml.text` | Stylized fonts, memes, decorative text | 0.85-0.90 | Free | ~200ms |
| Tesseract | `ocr.text` | Clean standard text, high contrast | Varies | Free | ~50ms |
---
## Cost Analysis
### Before Three-Tier System
100 images, all using Vision LLM:
```
100 images × $0.005/image = $0.50
Total time: 100 × 2s = 200 seconds
```
### After Three-Tier System
Route distribution (typical):
- 60 images → FAST route (Florence-2 only, free, ~100ms)
- 25 images → BALANCED route (Florence-2 + Tesseract, free, ~300ms)
- 10 images → QUALITY route (+ Vision LLM, $0.005, ~2s)
- 5 images → ANIMATED route (filmstrip, $0.002, ~2.5s)
```
Cost:
60 × $0 = $0
25 × $0 = $0
10 × $0.005 = $0.05
5 × $0.002 = $0.01
Total: $0.06
Time:
60 × 0.1s = 6s
25 × 0.3s = 7.5s
10 × 2s = 20s
5 × 2.5s = 12.5s
Total: 46 seconds
Savings:
Cost: 88% reduction ($0.50 → $0.06)
Time: 77% reduction (200s → 46s)
```
**The middle tier (Florence-2) handles 85% of images at zero cost.**
---
## Putting It All Together
Here's the full flow for a meme GIF with subtitles:
```
1. Load image: anchorman-not-even-mad.gif (93 frames)
2. IdentityWave (priority 10):
→ identity.frame_count = 93
→ identity.format = "gif"
→ identity.is_animated = true
3. TextLikelinessWave (priority 40, ~10ms):
→ Heuristic text detection: 15 regions in bottom 30%
→ Subtitle pattern: DETECTED
→ text.likeliness = 0.85
4. OcrWave (priority 50, ~60ms):
→ Run Tesseract OCR on first frame
→ ocr.text = "I'm not emn mad." (garbled)
→ ocr.confidence = 0.62
5. MlOcrWave (priority 51, ~180ms):
→ Tesseract confidence < 0.95, run Florence-2
→ Sample 10 frames (animated GIF)
→ Run Florence-2 on each frame (parallel)
→ Deduplicate: 10 results → 2 unique texts
→ ocr.ml.multiframe_text = "I'm not even mad.\nThat's amazing."
→ ocr.ml.confidence = 0.91
6. OcrQualityWave (priority 58, ~5ms):
→ Check Florence-2 result
→ Spell check: 6/6 words correct (100%)
→ ocr.quality.is_garbled = false
→ ocr.quality.escalation_required = false
7. VisionLlmWave (priority 80, SKIPPED):
→ No escalation required (Florence-2 succeeded)
Final output:
Text: "I'm not even mad.\nThat's amazing."
Source: ocr.ml.multiframe_text
Confidence: 0.91
Cost: $0 (local processing)
Time: ~250ms total (Tesseract + Florence-2)
```
If Florence-2 had failed (confidence < 0.5), the flow would continue:
```
6. OcrQualityWave:
→ Spell check: 2/6 words correct (33%)
→ ocr.quality.is_garbled = true
→ ocr.quality.escalation_required = true
7. VisionLlmWave:
→ Create text-only filmstrip (2 regions, 450×49)
→ Send to Vision LLM: "Extract all text from this strip"
→ vision.llm.text = "I'm not even mad.\nThat's amazing."
→ Confidence: 0.95
→ Cost: ~$0.002 (30× token reduction vs full frames)
→ Time: ~2.3s
```
---
## Configuration
The three-tier system is fully configurable:
```json
{
"DocSummarizer": {
"Ocr": {
"Tesseract": {
"Enabled": true,
"DataPath": "/usr/share/tesseract-ocr/4.00/tessdata",
"Languages": ["eng"],
"EarlyExitThreshold": 0.95
},
"Florence2": {
"Enabled": true,
"ModelPath": "models/florence2-base",
"ConfidenceThreshold": 0.85,
"MaxFrames": 10,
"DeduplicationMethod": "levenshtein",
"LevenshteinThreshold": 0.85
},
"Quality": {
"SpellCheckThreshold": 0.5,
"EscalationEnabled": true
}
},
"VisionLlm": {
"Enabled": true,
"Provider": "ollama",
"OllamaUrl": "http://localhost:11434",
"Model": "minicpm-v:8b",
"MaxRetries": 3,
"TimeoutSeconds": 30
},
"Filmstrip": {
"TextOnlyMode": true,
"SubtitleRegionPercent": 0.3,
"BrightPixelThreshold": 200,
"TextChangeThreshold": 0.05
},
"Routing": {
"FastRouteConfidence": 0.8,
"BalancedRouteConfidence": 0.5,
"TextDetectionEnabled": true
}
}
}
```
---
## Failure Modes
| Failure | Detection | Response |
|---------|-----------|----------|
| **Tesseract fails** | Confidence < 0.7 OR spell check < 0.5 | Escalate to Florence-2 |
| **Florence-2 fails** | Confidence < 0.5 OR spell check < 0.5 | Escalate to Vision LLM |
| **Vision LLM timeout** | Request exceeds 30s | Fall back to best available OCR result |
| **All tiers fail** | All results empty or garbled | Return empty string with confidence 0.0 |
| **API cost limit** | Daily budget exceeded | Disable Vision LLM, use Florence-2 only |
| **Model not available** | Florence-2/Vision LLM offline | Skip tier, continue to next |
Every failure is deterministic and logged with full provenance.
---
## Comparison to Other Approaches
### Traditional: Tesseract + Manual Fallback
```
For each image:
1. Run Tesseract
2. If looks wrong, manually fix or skip
Problems:
- No middle tier (binary: works or doesn't)
- Manual intervention required
- No cost optimization
```
### Cloud-First: Always Use Vision LLM
```
For each image:
1. Send to GPT-4o/Claude
2. Pay $0.005-0.01 per image
Problems:
- Expensive (85% of images could be free)
- Slow (network latency)
- Still hallucinates without constraints
```
### Three-Tier: Local-First with Smart Escalation
```
For each image:
1. OpenCV text detection (5-20ms, free)
2. Route to appropriate tier
3. Florence-2 handles 85% locally (200ms, free)
4. Vision LLM only for complex cases (2-5s, $0.001-0.01)
Benefits:
- 88% cost reduction
- 77% faster (most images process locally)
- Deterministic escalation (auditable)
- Filmstrip optimization (30× token reduction)
- Constrained by deterministic signals
```
---
## Conclusion
The three-tier OCR pipeline proves that **cost-aware routing** and **local-first processing** can dramatically improve both performance and economics without sacrificing quality.
Key insights:
1. **Florence-2 ONNX is the sweet spot**: Better than Tesseract for stylized fonts, faster and cheaper than Vision LLMs
2. **Text-only strips achieve 30× token reduction**: Extract bounding boxes, not full frames
3. **Routing is deterministic**: OpenCV detection + confidence thresholds, no guessing
4. **Escalation is auditable**: Every tier emits signals with provenance
5. **Failure is graceful**: Priority chain ensures fallback to best available source
The pattern scales: **local deterministic analysis → local ML model → cloud escalation**, each tier with known characteristics and cost trade-offs.
This is Constrained Fuzziness applied to OCR: deterministic signals (spell check, text detection) constrain probabilistic models (Florence-2, Vision LLM), and the final output aggregates sources by quality.
---
## Resources
### LucidRAG Documentation
- **[ImageSummarizer Library](https://github.com/scottgal/lucidrag/tree/main/src/Mostlylucid.DocSummarizer.Images)** - Source code
- **[Vision OCR Integration](https://github.com/scottgal/lucidrag/blob/main/src/Mostlylucid.DocSummarizer.Images/docs/vision-ocr-integration.md)** - Routing, filmstrips, token economics
- **[Architecture Guide](https://github.com/scottgal/lucidrag/blob/main/src/Mostlylucid.DocSummarizer.Images/docs/architecture.md)** - Waves, signals, escalation
- **[Pipeline Documentation](https://github.com/scottgal/lucidrag/blob/main/src/Mostlylucid.DocSummarizer.Images/docs/pipelines.md)** - Auto, balanced, quality routes
- **[Signals Reference](https://github.com/scottgal/lucidrag/blob/main/src/Mostlylucid.DocSummarizer.Images/docs/signals.md)** - Complete signal catalog
### CLI Tools
- **[ImageSummarizer CLI](https://github.com/scottgal/lucidrag/tree/main/src/Mostlylucid.ImageSummarizer.Cli)** - Command-line tool
- **[CLI README](https://github.com/scottgal/lucidrag/blob/main/src/Mostlylucid.ImageSummarizer.Cli/README.md)** - Usage and configuration
- **[Demo Images](https://github.com/scottgal/lucidrag/tree/main/src/Mostlylucid.ImageSummarizer.Cli/demo-images)** - Sample GIFs and frame strips
### Research Papers
- **[Florence-2 Paper](https://arxiv.org/abs/2311.06242)** - Microsoft's vision-language model
- **[EAST Text Detector](https://arxiv.org/abs/1704.03155)** - Efficient scene text detection
- **[CRAFT Paper](https://arxiv.org/abs/1904.01941)** - Character region awareness
- **[Real-ESRGAN](https://arxiv.org/abs/2107.10833)** - Practical super-resolution
- **[CLIP Paper](https://arxiv.org/abs/2103.00020)** - Learning transferable visual models
### Related Articles
- **[Part 4: Image Intelligence](/blog/constrained-fuzzy-image-intelligence)** - Wave architecture overview
- **[DocSummarizer](/blog/building-a-document-summarizer-with-rag)** - Document analysis pipeline
- **[DataSummarizer](/blog/datasummarizer-how-it-works)** - Data profiling approach
---
## The Series
| Part | Pattern | Focus |
|------|---------|-------|
| 1 | [Constrained Fuzziness](/blog/constrained-fuzziness-pattern) | Single component |
| 2 | [Constrained Fuzzy MoM](/blog/constrained-mom-mixture-of-models) | Multiple components |
| 3 | [Context Dragging](/blog/constrained-fuzzy-context-dragging) | Time / memory |
| 4 | [Image Intelligence](/blog/constrained-fuzzy-image-intelligence) | Wave architecture, patterns |
| **4.1** | **The Three-Tier OCR Pipeline (this article)** | **OCR, ONNX models, filmstrips** |
**Next**: Part 5 will show how ImageSummarizer, [DocSummarizer](/blog/building-a-document-summarizer-with-rag), and [DataSummarizer](/blog/datasummarizer-how-it-works) compose into multi-modal graph RAG with LucidRAG.
All parts follow the same invariant: **probabilistic components propose; deterministic systems persist**.