AI academy- Visual Pattern Recognition with CNNs: A Multi-Scene Case Study

This workshop turns the foundations of convolutional neural networks into a complete visual recognition case study.

The day begins by reframing images as structured data in which nearby pixels, local patterns, and spatial hierarchies matter. Participants revisit the mechanics of convolutional filters, receptive fields, pooling, nonlinear activations, and hierarchical feature learning, connecting each concept to what a model must learn in order to distinguish visually similar scenes.

The central case study uses a multi-class scene recognition problem containing visually diverse environments. Participants examine how image resizing, normalisation, class balance, train-validation splits, and augmentation affect model behaviour. Particular attention is given to confusion between related scene categories and to the difference between high validation accuracy and reliable recognition in less controlled images.

In the hands-on lab, participants implement and train a compact CNN in Keras/TensorFlow, monitor learning curves, inspect a confusion matrix, and explore intermediate activations to understand which visual cues the network has learned. They then test the model on altered or challenging images to identify failure modes.

Morning

Keynote Lecture: How CNNs Learn to See

Topics covered through the keynote

• Images as tensors and spatially structured data
• Convolutional filters, feature maps, and receptive fields
• Pooling, nonlinearities, and hierarchical feature learning
• CNN architectures for multi-class image recognition
• Data augmentation, regularisation, and overfitting
• Class imbalance, confusion, and evaluation beyond accuracy
• Visual ambiguity and dataset bias in scene recognition

Hands-on Lab: Multi-Scene Recognition with Keras

• Prepare and augment an image dataset
• Build and train a compact CNN
• Monitor training and validation behaviour
• Inspect confusion patterns and intermediate activations
• Test predictions on challenging scene variations
  
  

Afternoon

Group Case Study: Making Scene Recognition More Robust

Participants work in groups to diagnose model failures across different scenes and define:

• the most difficult visual classes and likely reasons for confusion
• data or augmentation strategies to improve generalisation
• architecture or training changes worth testing
• evaluation criteria for a more realistic deployment setting
  
  

Group Presentations and Model Critique

Short presentations (approximately 5 minutes per group)
Discussion of robustness, bias, and transfer to new visual environments

By the end of the course, participants will be able to:

• explain how CNNs learn hierarchical visual representations
• prepare image data for supervised visual recognition
• build and train a compact CNN using Keras/TensorFlow
• apply data augmentation and regularisation strategies
• interpret learning curves, confusion matrices, and feature activations
• diagnose common failure modes in multi-scene recognition
• propose practical strategies for improving visual model robustness
• 
  

One-day intensive workshop combining:

• case-study keynote lecture
• hands-on CNN implementation in Jupyter notebooks
• guided image preprocessing and augmentation
• group-based robustness challenge
• presentations and model critique