INFS3822 Chap.9 Unstructured Data and Image Inference
Unstructured Data and Image Inference
Text, images, audio and video must be transformed into structured representations before models can compare them. This chapter covers n-grams, named entities, sentiment features, convolutional feature maps and dimensionality reduction. It then traces a pretrained image-inference contract through resize, channel order, normalisation, batch shape, label order, confidence and human review.
What this chapter covers
- 01
Text feature extraction
- 02
Image feature mapping
- 03
Dimensionality reduction
- 04
Pretrained CNN inference
- 05
Resize and normalisation
- 06
Batch and label order
- 07
Confidence thresholds
- 08
Image clustering boundaries
Worked example · free
Interpret a three-class image prediction
- +1The largest probability is 0.67 in the surface-damage position.
- +1Because 0.67 is below the 0.75 automation threshold, the case cannot be rejected automatically.
- +1Route it as a surface-damage candidate for manual review and retain the image, scores and model version.
Key terms
- Convolutional network
- A neural architecture that learns spatial feature hierarchies from image-like inputs.
- Pretrained model
- A model whose parameters were learned on an earlier source dataset or task.
- Normalisation
- Transforming input values to the scale and distribution expected by the model.
- Confidence threshold
- A policy cutoff that routes lower-confidence outputs to review or another safe path.
Unstructured Data and Image Inference FAQ
Why does preprocessing need to match training?
A model learns patterns in a particular input representation. Changed channel order, scale, crop, resize or field mapping can move live inputs away from that representation and corrupt predictions. Document the contract and test reference examples whenever the deployment pipeline changes.
Can a confidence score be treated as certainty?
No. It is a model output under particular training and calibration conditions. Distribution shift, unfamiliar classes and preprocessing defects can make confidence unreliable. Use reviewed thresholds, low-confidence routes and labelled operational monitoring rather than treating a high number as proof.
Does an image cluster prove a defect category?
No. Clustering groups similarity in the chosen representation without supplied defect labels. Review representative images, assess stability and compare with verified outcomes. The cluster can support investigation or later supervised labelling, but it cannot establish defect status by itself.
What should a pipeline reference pack contain?
Keep ordinary, boundary and unfamiliar inputs with expected decoding, shape, channel order, scale, label mapping and final action. Run the pack after a change to preprocessing, model files or labels. A mismatch at an intermediate stage localises the defect and prevents premature retraining when the real problem is an inconsistent interface.
Assessment move
Memorise the inference contract in order: decode, colour channels, resize or crop, normalise, batch, predict, map labels, apply review policy. Debug that interface before changing model weights. For clustering, always write what representation created similarity and which external evidence validates the group. Build one reference image and record its expected shape, channel order, numeric range, label order and model output.
Walk it through training and deployment implementations and compare each intermediate representation. A changed crop, stretched aspect ratio, BGR input, missing batch dimension or different normalisation can reverse the meaning of a prediction without changing the model file. Keep the original input, transformed tensor, model version and label map together when investigating a failure.
Separate feature extraction from the downstream task. Text n-grams, named entities and sentiment features encode different relationships; CNN feature maps, pretrained embeddings, principal components and autoencoder codes create different notions of similarity. Ask what variation the representation preserves and what it discards.
When clustering extracted features, inspect representative members, stability and external outcomes before naming a group. An unsupervised cluster does not become a defect label by assertion. Practise probability interpretation with an ordered label list. Identify the maximum value, map its index to the documented class and apply the review threshold.
Then change the label order without changing the vector and explain why the action can reverse. Treat confidence as model output under particular conditions, not a guarantee. Use labelled operational samples to choose thresholds, report how many cases are deferred and monitor error by class and changing conditions such as lighting or product mix. When performance declines, audit the interface before retraining.
Test decoding, channels, dimensions, scaling, labels and camera conditions with known examples. If the pipeline is correct, investigate distribution shift and representation limits. End with a safe failure path for unfamiliar classes and low confidence: preserve evidence, route to human review and use verified labels to support later evaluation rather than forcing an automatic class.
Working through Unstructured Data and Image Inference in INFS3822? Sia is AskSia’s AI Computer Science tutor — ask any INFS3822 Unstructured Data and Image Inference question and get a clear, step-by-step explanation grounded in how INFS3822 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
