FloorMaterialAnalyzer/开发文档2

Material Analyzer Development Specification v2

Version: 2.0

Status: Development

Depends on: Material Analyzer Phase 1

1. Objective

The goal of Version 2 is not to introduce new downstream features such as Material QA or Texture Generation.

Instead, Version 2 focuses on upgrading the Material Analyzer into a production-grade offline analysis engine.

After completion, every flooring SKU should have a complete, stable, reproducible Material Asset that can serve as the Ground Truth across all future AI models.

2. Replace Placeholder Models

Current implementation uses deterministic placeholder vectors.

Replace them with real vision models.

2.1 DINOv2

Purpose

Extract texture-level visual embeddings.

Recommended model

facebook/dinov2-large

Output

1024-dim float32 vector

Store

embedding.bin

Metadata

{
    "model":"dinov2-large",
    "dimension":1024,
    "normalize":true
}

Applications

Texture similarity
Material similarity
Drift detection
SKU retrieval
2.2 CLIP

Purpose

Extract semantic visual embeddings.

Recommended

ViT-L/14

Output

768-dim float32 vector

Applications

Semantic similarity
Recommendation
Search
3. Vision LLM Semantic Analyzer

Replace rule-based semantic generation.

Recommended models

Priority

Gemini 3 Pro Image

GPT-4o Vision

Claude Vision

Input

Texture Image

Output Schema

{
    "description":"",
    "material_type":"",
    "surface_finish":"",
    "grain_type":"",
    "color_family":"",
    "visual_style":"",
    "variation":"",
    "gloss_level":"",
    "tags":[]
}

Important

LLM output should always be validated against JSON Schema.

Missing fields should be regenerated.

4. Feature Versioning

Every generated Material Asset must contain version information.

Example

{
    "analyzer_version":"2.0.0",

    "feature_schema":"2026.07",

    "generated_at":"ISO8601",

    "generator":"Material Analyzer"
}

Future schema changes must remain backward compatible.

5. Feature Validation

After extraction, automatically validate all features.

Example checks

Brightness ∈ [0,1]

Contrast ∈ [0,1]

Histogram length == 256

Embedding dimension == expected

Orientation ∈ [0,180]

Entropy > 0

Invalid assets should be regenerated.

6. Material Confidence

Every inferred feature should include a confidence score.

Example

{
    "stone_type":"Travertine",

    "confidence":0.91
}

This enables future QA weighting.

7. Canonical Material Statistics

Generate global statistics for every SKU.

Examples

Mean RGB

Median RGB

Color Variance

Texture Variance

Brightness Distribution

Dominant Orientation

Gradient Histogram

These statistics are independent of LLM output.

8. Asset Manifest

Every asset folder should contain

manifest.json

Example

{
    "sku":"67907_847",

    "files":[
        "preview.jpg",
        "thumbnail.jpg",
        "material.json",
        "embedding.bin",
        "histogram.json"
    ],

    "status":"complete"
}

This simplifies integrity checking.

9. Batch Processing

Improve offline processing.

Support

Incremental Update

Resume

Retry

Multi-thread

Progress Bar

Logging

Failure Report

Target

3500+ SKU

One-click processing

Resume after interruption
10. Analyzer Benchmark

Generate benchmark reports after every batch.

Example

Total SKU

Processed

Skipped

Failed

Average Time

Model Version

Feature Version

Image Resolution Distribution

Output

benchmark.json
11. Plugin Architecture

Future feature extractors should be pluggable.

Directory

analyzers/

    color/

    texture/

    semantic/

    embedding/

    adapters/

    plugins/

Every extractor should implement

Analyze(image) -> Feature

No module should depend directly on another module.

12. Future Compatibility

Analyzer output must remain stable for

Material QA
Canonical Texture Generator
Tile Detection
Pattern Detection
Similarity Search
AI Prompt Builder
Product Recommendation

Do not tightly couple Analyzer with any downstream module.

13. Development Priority

Phase 1 (Completed)

Universal Features
Texture Features
Rule-based Semantic
Placeholder Embedding

Phase 2 (Current)

DINOv2
CLIP
Vision LLM
Versioning
Validation
Confidence
Manifest

Phase 3

Canonical Texture Generator
Tile Detection
Image Quilting
Pattern Detector

Phase 4

Material QA
Drift Detection
Auto Retry