Machine Learning Project: Image Classification Syllabus

Project: Image Classification syllabus — full chapter and topic list

Expand any chapter to see its topics and sub-topics. This is the whole examinable outline for Project: Image Classification in Machine Learning, not a summary of it.

1. 1

   Define the Project

   2 topics

   • Identify the Problem
     • Specify the Type of Images
     • Define the Objective
   • Gather Requirements
     • Determine Hardware and Software Needs
     • Identify Data Sources
2. 2

   Collect and Prepare Data

   3 topics

   • Collect Data
     • Download or Scrape Images
     • Label Images
   • Preprocess Data
     • Resize Images
     • Normalize Pixel Values
     • Augment Data
   • Split Data
     • Create Training Set
     • Create Validation Set
     • Create Test Set
3. 3

   Build the Model

   2 topics

   • Choose a Model Architecture
     • Select a Pre-trained Model or Build from Scratch
   • Compile the Model
     • Define Loss Function
     • Select Optimizer
     • Specify Metrics
4. 4

   Train the Model

   2 topics

   • Set Training Parameters
     • Define Batch Size
     • Set Number of Epochs
   • Monitor Training
     • Track Training and Validation Loss
     • Adjust Learning Rate if Necessary
5. 5

   Evaluate the Model

   2 topics

   • Test on Validation Set
     • Calculate Accuracy
     • Generate Confusion Matrix
   • Test on Test Set
     • Evaluate Performance Metrics
6. 6

   Deploy the Model

   2 topics

   • Choose Deployment Method
     • Deploy as a Web Service
     • Deploy on Mobile Device
   • Monitor and Maintain
     • Track Model Performance
     • Update Model as Needed

Project: Image Classification flashcards for Machine Learning

23 of 50 cards from the Project: Image Classification deck — real questions with worked answers.

1. What is the core goal of an image classification project?

   To build a model that assigns one or more predefined category labels to an input image, mapping pixel data to a discrete class.

2. In the 'Identify the Problem' stage, what distinguishes binary, multi-class, and multi-label classification?

   Binary = 2 mutually exclusive classes; multi-class = one label chosen from more than 2 mutually exclusive classes; multi-label = an image can belong to several classes simultaneously.

3. Why is defining success metrics part of identifying the problem?

   Because the metric (e.g., accuracy, precision/recall, F1) determines how you train, tune, and judge the model; an undefined target makes the project impossible to evaluate.

4. What is the difference between image classification and object detection?

   Classification assigns a label to the whole image; object detection also localizes objects with bounding boxes and classifies each one.

5. During 'Gather Requirements', what are the key categories of requirements to capture?

   Functional (what classes/accuracy), data (sources, volume, labeling), performance (latency, throughput), hardware/deployment constraints, and legal/privacy requirements.

6. What is a latency requirement and why does it matter for model choice?

   Latency is the maximum allowable inference time per image; tight latency budgets favor smaller, faster architectures (e.g., MobileNet) over large ones (e.g., ResNet-152).

7. When collecting data, what does 'class balance' mean and why is it important?

   Class balance is having roughly comparable numbers of examples per class; imbalance biases the model toward majority classes and inflates accuracy while hiding poor minority-class performance.

8. What is data leakage in the context of collecting and splitting image data?

   When information from outside the training set (e.g., near-duplicate images, or test images) leaks into training, producing overly optimistic and unrealistic performance estimates.

9. Why must collected image data be representative of deployment conditions?

   Because a model only generalizes to distributions similar to its training data; mismatched lighting, angles, or devices cause domain shift and degraded real-world accuracy.

10. What is image annotation/labeling, and what quality issue commonly arises?

    Labeling is assigning ground-truth class labels to images; the common issue is label noise (incorrect or inconsistent labels), which limits achievable accuracy.

11. In preprocessing, what is image resizing and why is it required?

    Resizing converts all images to a fixed spatial size (e.g., 224x224) because most networks require a uniform input shape for batching and matrix operations.

12. Write the standard min-max normalization that scales pixel values from $[0,255]$ to $[0,1]$.

    $$x' = \frac{x}{255}$$

13. Write the z-score (standardization) formula used to normalize image channels with mean $\mu$ and standard deviation $\sigma$.

    $$x' = \frac{x - \mu}{\sigma}$$

14. What is data augmentation and name four common image augmentations?

    Synthetically expanding training data by transforming images; examples: random horizontal flip, rotation, random crop, brightness/contrast jitter, and zoom.

15. Why does data augmentation help reduce overfitting?

    It increases effective dataset diversity, exposing the model to plausible variations so it learns invariant features instead of memorizing specific pixels.

16. What is one-hot encoding of class labels and why is it used with categorical cross-entropy?

    Representing a label as a vector with a 1 at the true class index and 0 elsewhere; categorical cross-entropy compares this target vector against the predicted probability distribution.

17. What is a typical train/validation/test split ratio, and what is each subset used for?

    Commonly 70/15/15 or 80/10/10. Train fits parameters, validation tunes hyperparameters and monitors generalization, test gives a final unbiased performance estimate.

18. Why must the test set remain untouched until the very end?

    To provide an unbiased estimate of generalization; using it during development causes information leakage and over-optimistic results.

19. What is stratified splitting and when should you use it?

    Splitting so each subset preserves the original class proportions; use it for imbalanced datasets to ensure all classes appear in train, validation, and test.

20. What is k-fold cross-validation and what does it estimate?

    The data is split into $k$ folds; the model trains on $k-1$ folds and validates on the remaining one, rotating $k$ times. It estimates generalization more robustly by averaging across folds.

21. What is a Convolutional Neural Network (CNN) and why is it suited to images?

    A network using convolutional layers that apply learnable filters across spatial regions; it exploits local connectivity and weight sharing to capture spatial patterns efficiently.

22. For an input of size $W$, filter size $F$, padding $P$, and stride $S$, give the output spatial dimension of a convolution layer.

    $$O = \frac{W - F + 2P}{S} + 1$$

23. What is the role of a pooling layer (e.g., max pooling) in a CNN?

    It downsamples feature maps to reduce spatial dimensions and computation while providing translation invariance; max pooling keeps the maximum value in each window.

See more Project: Image Classification flashcards →

Planning Project: Image Classification for Machine Learning

Project: Image Classification (Machine Learning) FAQ