Potential model and feature additions

[EHxuban11]

Potential additions to consider, scoped to existing tasks and infrastructure:

• Table Transformer: table detection. Detect tables and rotated tables in document images, with pretrained loading, fine-tuning and ONNX export parity.
• Table Transformer: structure recognition. Detect rows, columns, headers and spanning cells within table crops. Return labeled boxes; leave PDF parsing, OCR and spreadsheet reconstruction for separate work.
• Swin classifier fine-tuning. Extend the existing inference-only family with a classification trainer, starting with Swin-T. Support replacing the head, resuming training and validation, with convergence evidence.
• DenseNet-121 classification. Add pretrained inference, fine-tuning and ONNX export using the existing classification contracts. Start with one size.
• Input-equivalence coverage for YOLO9 and RF-DETR. Extend existing tests to compare path, PIL and NumPy inputs, individual versus batched predictions, and mixed image sizes. Verify image count, order, labels and original-image coordinates.
• Checkpoint-specific license information in the model catalogue. Expose separate code and weight license links and restrictions, using the existing notices without changing checkpoint schemas.
• Executable flagship workflow examples. Maintain YOLO9 and RF-DETR examples covering loading, prediction, training, validation, export and reload. Reuse the examples in documentation and run them as smoke checks.
• Optional intermediate-feature inspection. Retrieve selected layer outputs and shapes for YOLO9 and RF-DETR without changing normal prediction behavior or the Results contract.

Use permissively licensed original sources for new ports, verify checkpoint redistribution terms, and keep model-specific changes isolated. These are candidates for separate, bounded contributions, not a proposal to rewrite shared architecture.