Recognition Confidence Scores

[sneakybatman]

Hey team,

When computing word-level confidence in recognition, the current implementation appears to take the arithmetic mean of per-character confidences. This can hide single-character mistakes: one very low-confidence (and incorrect) character can be washed out by many high-confidence characters, yielding an over-optimistic word score.

I propose making the aggregation strategy configurable, with options such as:

• mean (current behavior, for backward compatibility)

• min (pessimistic; flags single-char issues)

• geomean (product^(1/N); reflects sequence probability)

• avg_log_prob (exp(mean(log p_i)); numerically stable version of geomean)

• (optional) harmonic or trimmed_mean (robust alternatives)

Downstream systems often gate automation vs. human review using word-level confidence. For IDs, names, dates, etc., a single wrong character can invalidate the field. An aggregator that better penalizes single-char errors yields more reliable thresholds and fewer false passes.

Example

Per-character confidences for a 4-char word: [0.99, 0.98, 0.12, 0.99]

Aggregator	Result
Arithmetic mean	0.770
Geometric mean	0.583
Avg log-prob	0.583
Harmonic mean	0.352
Min	0.120

The arithmetic mean remains high (0.77) despite one very low-confidence character.

[e68026063-oss]

The same can be said about box_score for the detection models.

[e68026063-oss]

That said if you look at the implementations, the confidence score for recognition models depends on the model type. For example, if it is an attention based model (ex: parseq, vitstr), then it is indeed computed as a mean. But if you look at the others (crnn, master, ...), then the confidence of the sequence is the min of conf from all chars. What is the name of the reco model that you used ?