City University of Hong Kong · FACULTY OF INFORMATION TECHNOLOGY

IS6400 Chap.9 Image Mining and Recognition

- one subject, every graph, every model, every mark
6 Chapters2-page Bible
Our own words - no uploaded lecturer files
Updated for this semester
Chapter 9 of 10 · IS6400

Image Mining and Recognition

Pixels become features and scores

Image models learn local filters, combine them into feature maps, reduce spatial detail and produce class scores.

Strong performance can come from background, camera or annotation shortcuts rather than the object of interest, so the capture process belongs inside the model audit.

Counts reveal the operational queue

Begin evaluation with true and false positive and negative counts at the chosen threshold. Precision, recall and accuracy answer different questions.

Split by capture session, device, site or time when similar frames could otherwise appear in training and test.

Treat capture context as data

Record device, site, session, angle, lighting, operator and annotation status when available. A defect model can learn a table, date stamp or camera rather than the product.

Random image splits are especially misleading when near-duplicate frames from one capture run enter both training and test. Hold out sessions, devices, sites or later time periods according to the intended deployment and keep independently captured images for the strongest check.

Design challenge sets around shortcuts

Change background, crop, blur, occlusion and lighting while preserving the object label.

Compare error by product family and capture condition, not only across the full set. Heat maps can reveal attention to an obvious shortcut but do not prove the network's reasoning. Performance on deliberately varied and independent images is stronger evidence.

Document missing contexts so a high average score is not interpreted as coverage the evaluation never supplied.

Connect the threshold to review work

Construct true positives, false positives, false negatives and true negatives before converting them to rates. Recall divides detected defects by real defects; precision divides true defects by all flagged images.

Translate false positives into review minutes and false negatives into escaped-defect cost, then verify the queue against staffing. Recheck the chosen threshold by context and create an escalation route for images outside the represented capture range.

In this chapter

What this chapter covers

  • 01

    Pixel grids and local filters

  • 02

    Feature maps and pooling

  • 03

    Capture-context shortcuts

  • 04

    Session-level data splits

  • 05

    Confusion counts

  • 06

    Threshold cost and review capacity

Worked example · free

Measure a defect-recognition queue

Q [8 marks]. AskSia-authored practice. The listed marks are AskSia study allocations, not a University rubric. Of 100 items, 20 are defective. A model flags 25 items, including 16 true defects. Compute recall, precision and accuracy.
The arrows reduce spatial detail while increasing abstraction.The arrows reduce spatial detail while increasing abstraction. Local pixel patterns become feature maps, pooled evidence and finally class probabilities that still require threshold and error review.Pixels become local features, pooled evidence and a class scoreimage gridfeature mapspoolingclass probabilities
The arrows reduce spatial detail while increasing abstraction. Local pixel patterns become feature maps, pooled evidence and finally class probabilities that still require threshold and error review.
  • 3Construct all four confusion counts. The mark allocation here is not an official University marking scheme.
  • 3Compute recall and precision.
  • 2Compute accuracy and explain what it hides.
True positives are 16, false positives 9, false negatives 4 and true negatives 71. Recall is 16/20 = 0.80. Precision is 16/25 = 0.64. Accuracy is (16 + 71)/100 = 0.87. The accuracy headline hides four defective items released and nine sound items sent to review. Select the threshold using the cost of a missed defect and available review capacity, then retest on independently captured images.
Sia tip — Write the four counts before calculating any rate; the queue and the harmful misses are visible before the percentages are.
Glossary

Key terms

Convolutional Filter
A learned local pattern detector applied across an image grid.
Feature Map
The spatial response produced when a filter is applied across an image.
Pooling
A reduction step summarising nearby feature responses and decreasing spatial detail.
Capture Context
The camera, site, lighting, background and session conditions under which an image was produced.
Shortcut Feature
An incidental pattern that predicts labels in one data set without representing the intended object.
FAQ

Image Mining and Recognition FAQ

Why split images by session instead of randomly?

Near-duplicate frames from one session can otherwise appear on both sides, inflating performance. A session, device, site or time split better tests whether the model transfers to new capture conditions.

What does a saliency map prove?

It can reveal obvious attention to a background or stamp, but it does not prove the model's reasoning. Stronger evidence comes from controlled context changes and independently captured test images.

How should a recognition threshold be selected?

Compare confusion counts and error costs at thresholds the review team can operate. A costly missed defect may justify more false alarms, but the resulting queue still has to fit capacity.

Study strategy

Exam move

Draw the image pipeline from grid to feature maps, pooling and class score. Create a confusion matrix from counts and compute recall, precision and accuracy with denominators named. Then design a challenge set that changes background, lighting, angle and device while preserving the object. Finish by stating the costly error, chosen threshold, review capacity and capture context not yet represented in testing.

Make a capture data sheet with session, site, device, operator and annotation status, and use it to design a group-aware split. Contrast that split with a random image split containing near-duplicate frames. Recompute the queue after moving the threshold in both directions, translating false alarms into review minutes and missed defects into escaped cost.

Add independently captured images and report performance by product family and context, including any condition too rare to support a reliable claim. Practise separating a saliency clue from validation evidence: only controlled context changes and independent captures show whether performance survives a suspected shortcut. End by specifying the escalation route for an unfamiliar camera or image condition.

A+Everything unlocked
Unlocks this Bible + all 31 of your subjects at City University of Hong Kong, plus 1,000+ more Bibles across the AskSia library.
Sia - your IS6400 tutor, unlimited, worked the way the exam marks it
The full 2-page Bible + practice bank with worked solutions
Chrome extension - sync your LMS so Sia knows your deadlines
Bilingual EN / Chinese on every Bible and every Sia answer
$0.99 Trial
30-day money-back · cancel in one tap · how it works