(Fixes PaddlePaddle/PaddleOCR#18064) fix: guard against KMeans n_clusters=0 crash in table recognition pipeline

[Mane2305]

(Fixes PaddlePaddle/PaddleOCR#18064)

Problem

When using PPStructureV3 with engine="transformers" and
wireless_table_structure_recognition_model_name="SLANeXt_wireless",
the pipeline crashes with:

sklearn.utils._param_validation.InvalidParameterError:
The 'n_clusters' parameter of KMeans must be an int in the
range [1, inf). Got 0 instead.

Related issue: #5140

Root Cause

SLANeXt does not support resampling in the transformers engine
(logged as a warning), which causes zero cell detections.
This zero value propagates to KMeans as n_clusters=0 → crash.

Two code paths trigger this:

1. N = html_pred_boxes_nums - len(cells_det_results) can be <= 0
2. combine_rectangles() passes N directly to KMeans without validation

Fix

1. Added early return guard in combine_rectangles() when N <= 0 or
   rectangles list is empty
2. Added N <= 0 guard before calling combine_rectangles() in
   cells_det_results_reprocessing()

Testing

Tested locally with paddleocr==3.5.0, transformers==5.9.0 on Windows.
Pipeline now completes successfully instead of crashing.

Before fix:
[transformers] Resampling is not supported in SLANeXt
❌ InvalidParameterError: n_clusters=0

After fix:
[transformers] Resampling is not supported in SLANeXt
✅ Prediction done

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

All committers have signed the CLA.

[Mane2305]

Local Testing Proof

Before fix (crash):

After fix (working):

Tested on:

• OS: Windows 11
• paddleocr: 3.5.2
• transformers: 5.9.0
• torch: 2.8.0
• engine: transformers
• model: SLANeXt_wireless