Must every stage use the same model type?

No. Different tests or model families can be combined.

Why use a high-recall first stage?

It reduces the chance of discarding true targets before more detailed analysis.

Can a cascade include human review?

Yes. A workflow may route selected cases to people, although that adds operational considerations beyond classifier computation.

AI Glossary / AI and machine learning / Intermediate

Cascading classifiers

Definition

Cascading classifiers are classification stages arranged so that an example's result at one stage determines whether it proceeds to another. In a common detection cascade, inexpensive early stages reject obvious negatives and later stages examine the remaining candidates more closely.

What it means

Many detection tasks contain far more irrelevant candidates than useful ones. An image may contain thousands of possible windows, while only a few contain the object being sought. Running the most expensive classifier on every window can waste computation.

A cascade spends little effort on easy rejections and more effort on uncertain survivors. The Viola - Jones object detector is an influential example of this approach. Other systems can use cascades with different models or routing rules, so “cascade” describes the arrangement as much as the individual classifier.

An early decision can be difficult to recover from. In a reject-only cascade, a true target discarded by the first stage never reaches the later, more capable stage. The design therefore balances computational savings with the risk of premature rejection.

How it works

Each stage receives only the candidates that passed previous stages. It applies a test or classifier and either rejects the candidate or allows further processing. Early stages are often tuned for high sensitivity, accepting some false positives so that true targets are rarely lost.

Later stages are trained and evaluated with awareness of the harder examples that survive. Their input distribution differs from the original population. A stage's apparent accuracy on random examples may therefore say little about its performance inside the cascade.

For a strict sequence where all stages must pass, overall target retention is the product of the conditional retention rates. The same principle applies to conditional false-positive pass rates. Computational cost depends on how many candidates reach each stage and the cost of that stage. The complete system must be evaluated end to end, including thresholds, routing and any fallback.

Three practical examples

1. An image detector quickly rejects windows that clearly lack a target and reserves detailed classification for plausible windows.

2. A document-processing pipeline first identifies likely forms using a cheap test, then applies a more costly model to the surviving pages. It measures missed forms as well as processing time.

3. A content-analysis service uses a lightweight model to route uncertain cases to a larger model. Its routing arrangement is a cascade, although its acceptance rules may differ from a strict object-detection cascade.

Think of it this way

Think of several inspection desks along a path. The first checks obvious requirements quickly, while later desks handle the cases needing closer attention. The limit is that a computational rejection can silently remove an item from further consideration; the later desks cannot repair a decision about an item they never receive.

Worked example

Suppose 10,000 candidates enter a two-stage detection cascade. Stage one costs 1 unit per candidate and passes 20%, so its cost is 10,000 units. Stage two costs 10 units for each of the 2,000 survivors, adding 20,000 units. Total cost is 30,000 units, compared with 100,000 for applying stage two to every candidate.

For true targets, suppose stage one retains 98%, and stage two retains 95% of the true targets that reach it. Overall retention is 0.98 × 0.95 = 93.1%.

The cost saving therefore comes with a quality question. The calculation does not establish that the resulting miss rate is acceptable for the intended use.

Illustrative case study

A fictional archive digitization team uses a fast classifier to identify pages likely to contain tables. Only those pages receive a slower extraction model. The initial timing result looks attractive, but a review finds that faintly printed tables are often rejected early.

The team changes the first-stage threshold and routes low-quality scans to a separate review path. It evaluates extraction coverage across scan-quality groups and records the additional processing cost.

The case shows why speed cannot be assessed in isolation. A cascade can make a downstream model appear efficient by preventing difficult but important examples from ever reaching it.

Common misconceptions and mistakes

Later stages fix every earlier mistake. They cannot fix rejected cases they never see unless a separate recovery path exists.
Stage accuracies can simply be averaged. Overall performance depends on conditional decisions and routing.
A cascade always improves speed. Savings depend on stage costs, pass rates and implementation overhead.

Business use

Cascades can help control latency and computing costs in high-volume classification or detection. Begin by identifying which errors are acceptable at each stage and which require a review route.

Measure end-to-end recall, false positives, latency and cost on representative workloads. Track the fraction reaching each stage because a change in input data can erase expected savings. Evaluate important subgroups and difficult conditions separately. Maintain visibility into early rejections so that a fast first stage does not become an unexamined source of lost cases.

Check your understanding

A first stage retains 90% of true targets, and a second retains 90% of those survivors. Is total retention 90%?

Answer: No. In this strict two-stage cascade it is 0.90 × 0.90 = 81%.

Related terms

Recall
False positive
Early exit
classifier
object detection
Viola–Jones detector
model routing
computational cost

Learn next

Read recall and precision, then diagram a two-stage workflow. Calculate both the number of candidates reaching each stage and the true targets lost before the final decision.

Recall

Sources

OpenCV, official Cascade Classifier tutorial, including the Viola–Jones approach: https://docs.opencv.org/4.x/db/d28/tutorial_cascade_classifier.html. Viola and Jones, Rapid Object Detection Using a Boosted Cascade of Simple Features (2001), DOI: https://doi.org/10.1109/CVPR.2001.990517. All processing costs, rates and archive scenarios above are invented.

Previous: Cascaded finite-state transducers Next: Case agreement