Advanced Computer Vision & Image Recognition

Course outline and learning expectations

This is a tutor-led course. The outline shows what your tutor will cover; teaching materials and examinations are provided directly to enrolled students.

What you will learn

• Explain modern object detection pipelines and the core ideas behind YOLO
• Collect, label and split an image dataset for a defined detection task
• Train, validate and evaluate a custom object detector using precision, recall, IoU and mAP
• Apply semantic and instance segmentation concepts to image-analysis problems
• Evaluate object and face recognition capabilities while assessing bias, consent, privacy and ethical limits
• Create a real-time camera inference pipeline for Android or another edge device
• Quantise and export a model with LiteRT/TensorFlow Lite and compare accuracy, latency and model size
• Integrate preprocessing, inference, post-processing and user-interface components into an end-to-end vision app
• Deliver and document a deployable coffee cherry ripeness detector prototype for a basic Android phone

Modules

1. 1

   Object detection architectures and YOLO basics

   Understand detection pipelines and the foundations of YOLO-based object detection

   Computer vision pipelineBounding boxes and class labelsYOLO architecture basicsConfidence scores and non-maximum suppressionIntersection over UnionPrecision, recall and mean average precision
2. 2

   Building a custom object detector

   Create a task-specific detector from data collection through evaluation

   Problem definitionResponsible image collectionAnnotation formats and toolsTraining, validation and test splitsData augmentationModel trainingError analysis and evaluation
3. 3

   Image segmentation concepts

   Use pixel-level methods to separate objects and regions within images

   Semantic segmentationInstance segmentationSegmentation masksDataset annotationIoU and Dice metricsPractical segmentation workflow
4. 4

   Face and object recognition: capabilities and ethical limits

   Examine recognition methods, reliability limits and responsible deployment in Uganda and Africa

   Feature embeddingsSimilarity and decision thresholdsFalse matches and false rejectionsDataset biasConsent and lawful data usePrivacy and securityWhen not to deploy recognition
5. 5

   Real-time vision on mobile and edge devices

   Build an efficient camera-to-inference pipeline that can operate with limited connectivity

   Camera inputImage preprocessingOn-device inferencePost-processingFrames per second and latencyOffline operationDevice testing
6. 6

   Optimising models for low-power devices

   Reduce model size and latency while measuring the effect on accuracy

   Model profilingPost-training quantisationInteger and float quantisationLiteRT/TensorFlow Lite exportHardware acceleration and delegatesAccuracy-latency-size trade-offs
7. 7

   Building a vision app end-to-end

   Integrate the model, application logic and user interface into a testable mobile solution

   Application architectureCamera integrationModel loadingInference pipelineResults displayError handlingField testing and documentation
8. 8

   Capstone: deployable vision application

   Develop a coffee cherry ripeness detector prototype that runs on a basic Android phone

   Define ripeness classesCollect and annotate representative imagesTrain and evaluate the detectorOptimise the model for mobile useDeploy to AndroidTest under varied lighting and backgroundsDocument limitations and responsible use