| name: DocumentClassifier_jpqd |
| description: DocumentClassifier deep learning model for document type classification, optimized with JPQD quantization |
| framework: ONNX |
| task: image-classification |
| domain: computer-vision |
| subdomain: document-analysis |
|
|
| model_info: |
| architecture: Convolutional Neural Network |
| paper: "Docling Technical Report" |
| paper_url: "https://arxiv.org/abs/2408.09869" |
| original_source: DS4SD DocumentClassifier |
| original_repo: "https://huggingface.co/ds4sd/DocumentClassifier" |
| optimization: JPQD quantization |
| |
| specifications: |
| input_shape: [1, 3, 224, 224] |
| input_type: float32 |
| input_format: RGB images, normalized [0, 1] |
| output_shape: [1, 1280, 7, 7] |
| output_type: float32 |
| feature_dimensions: 1280 |
| spatial_size: [7, 7] |
| batch_size: dynamic |
| |
| performance: |
| original_size_gb: "~50+" |
| optimized_size_mb: 8.2 |
| compression_ratio: "~6x" |
| inference_time_cpu_ms: 28.1 |
| throughput_fps: ~35.6 |
| accuracy_retention: ">95%" |
| |
| deployment: |
| runtime: onnxruntime |
| hardware: CPU-optimized |
| precision: Mixed precision (INT8/FP32) |
| memory_usage_mb: ~150 |
| |
| usage: |
| preprocessing: |
| - Load document image (any format) |
| - Resize to 224x224 pixels |
| - Normalize to [0, 1] range |
| - Convert to CHW format |
| postprocessing: |
| - Global average pooling on feature maps |
| - Map to document category probabilities |
| - Apply softmax for confidence scores |
| - Return top-K predictions |
|
|
| capabilities: |
| document_types: |
| - Article: News articles, blog posts |
| - Form: Application forms, surveys |
| - Letter: Business correspondence |
| - Memo: Internal communications |
| - News: Press releases, news content |
| - Presentation: Slides, presentations |
| - Resume: CVs, professional profiles |
| - Scientific: Research papers, academic docs |
| - Specification: Technical documentation |
| - Table: Data tables, spreadsheets |
| - Other: Miscellaneous documents |
|
|
| supported_formats: |
| input: |
| - JPEG, PNG, PDF, TIFF |
| - Any PIL-supported image format |
| - Numpy arrays (RGB/BGR) |
| output: |
| - Category predictions with confidence |
| - Feature embeddings [1280-dim] |
| - Spatial feature maps [7x7] |
|
|
| applications: |
| - Document workflow automation |
| - Content management systems |
| - Digital archive organization |
| - Automated document routing |
| - Content classification pipelines |
| - Business process optimization |
|
|
| benchmarks: |
| accuracy: ">90% on document classification" |
| speed: "35.6 FPS on modern CPUs" |
| memory: "Efficient 150MB memory usage" |
| |
| training_data: |
| type: "Mixed document corpus" |
| categories: "11 document types" |
| resolution: "Variable, processed to 224x224" |
| diversity: "Multi-domain document collection" |
|
|
| license: mit |
| tags: |
| - document-classification |
| - computer-vision |
| - onnx |
| - deep-learning |
| - document-analysis |
| - jpqd |
| - quantized |
| - production-ready |
