Deep Learning and Computer Vision: Neural Networks, CNNs, Object Detection and Deployment

Programme overview

Introduction:

Deep learning and computer vision projects often stall between a promising notebook and a model that works on real plant images: training diverges, accuracy collapses on new lighting, detectors miss small defects and nobody can run the model on edge hardware. This Core Concept course builds code-first capability in PyTorch, from neural network mechanics and training discipline to convolutional networks, transfer learning, object detection, segmentation, video analytics and deployment. Every session is a lab, and participants finish with an Industrial Vision Model Evaluation Pack built on an industrial image dataset.

Course Objectives:

• Build and train neural networks in PyTorch, selecting activation functions, loss functions and optimisers suited to the task
• Diagnose underfitting and overfitting and control them with regularisation, dropout, normalisation and structured hyperparameter search
• Adapt pretrained convolutional and transformer backbones to small domain image datasets through transfer learning and fine-tuning
• Train and evaluate object detection and segmentation models, reporting IoU, mean average precision and error cases
• Export, compress and benchmark vision models for edge inference and track experiments, drift and model versions
• Deliver an Industrial Vision Model Evaluation Pack with trained model, evaluation report, deployment package and model card

Target Audience:

• Data science staff who build predictive models and now need to work with image and video data
• Machine learning engineering staff responsible for training, packaging and serving neural network models
• Technical analysts who prototype automation and analytics solutions in Python
• Automation and digital engineering staff who integrate camera systems with plant or field software
• Research and development staff who evaluate deep learning methods for new products and processes

Course Outline:

Day 1: Neural Network Foundations and the PyTorch Workflow

• Perceptron, Multilayer Perceptron and Decision Boundary Intuition
• PyTorch Tensors, GPU Devices and Autograd Computational Graphs
• Forward Pass and Backpropagation Through the Chain Rule
• Activation Functions: Sigmoid, Tanh, ReLU and GELU Behaviour
• Loss Functions: Mean Squared Error, Cross-Entropy and Focal Loss

Day 2: Optimisers and the Training Loop

• Stochastic Gradient Descent, Momentum, RMSprop and Adam Optimisers
• Learning-Rate Schedules: Step Decay, Cosine Annealing and Warm-Up
• Weight Initialisation Schemes and Vanishing or Exploding Gradients
• PyTorch Dataset, DataLoader and Mini-Batch Training Loop Build
• Training and Validation Curves with TensorBoard Logging

Day 3: Regularisation, Generalisation and Hyperparameter Search

• L1 and L2 Weight Decay with Early Stopping Callbacks
• Dropout Layers and Their Behaviour at Training Versus Inference Time
• Batch Normalisation and Layer Normalisation Placement
• Hyperparameter Search with Random Search and Optuna Trials
• Reproducibility Controls: Random Seeds, Deterministic Kernels and Config Files

Day 4: Convolutional Neural Networks and Transfer Learning

• Convolution, Stride, Padding, Pooling and Receptive Field Arithmetic
• LeNet-5, AlexNet, VGG and ResNet Residual Block Design
• torchvision Pretrained Backbones for Feature Extraction
• Fine-Tuning Strategies: Frozen Layers, Discriminative Learning Rates and Gradual Unfreezing
• Grad-CAM Saliency Maps for Inspecting CNN Decisions

Day 5: Sequence Models, Transformers and Week-One Integration Lab

• Recurrent Networks, LSTM and GRU Cells for Sensor Time Series
• Self-Attention, Multi-Head Attention and Positional Encoding
• Vision Transformer Patch Embedding Compared with CNN Backbones
• Keras Sequential and Functional API Equivalents of the PyTorch Models
• Guided Lab: Image Classifier Built, Tuned and Explained End to End

Day 6: Image Pipelines, Augmentation and Annotation

• Image Data Pipeline Design: Decoding, Resizing, Normalisation and Caching
• Albumentations Geometric, Photometric and Cutout Transforms
• Mixup and CutMix for Small Industrial Image Sets
• Annotation Formats: COCO JSON, Pascal VOC XML and YOLO Text Labels
• Class Imbalance Remedies: Weighted Sampling and Synthetic Minority Images

Day 7: Object Detection with Single-Stage and Two-Stage Detectors

• Anchor Boxes, Anchor-Free Heads and Non-Maximum Suppression
• YOLO-Family Single-Stage Detector Training with Ultralytics
• Faster R-CNN Two-Stage Detection with torchvision
• Intersection over Union and Mean Average Precision Calculation
• Small-Object and Occlusion Error Analysis on Detector Outputs

Day 8: Segmentation, Tracking and Video Analytics

• U-Net Encoder-Decoder Semantic Segmentation
• Mask R-CNN Instance Segmentation and Polygon Masks
• Dice Coefficient and Mean IoU Segmentation Metrics
• Multi-Object Tracking with SORT and ByteTrack Association
• Video Stream Ingestion with OpenCV and Frame-Sampling Strategies

Day 9: Deployment, Edge Inference, MLOps and Responsible AI

• ONNX Export and ONNX Runtime Inference Benchmarking
• Post-Training Quantisation, Pruning and TensorRT Optimisation for Edge Devices
• Experiment Tracking and Model Registry with MLflow
• Data Drift Monitoring and Retraining Triggers for Deployed Vision Models
• Responsible AI Checks: Dataset Bias Audit, Privacy Masking and Model Cards

Day 10: Capstone: Industrial Vision Model Build and Evaluation

• Industrial Dataset Brief: Surface Defect, Component Counting or Corrosion Detection
• Baseline Training Run and Augmentation Plan Execution
• Evaluation Report with Confusion Matrix, mAP and Latency Measurements
• Deployment Package: ONNX Model, Inference Script and Model Card
• Capstone Presentation and Technical Peer Code Review

Skills You Will Gain:

• Neural Network Implementation
• Gradient-Based Optimisation
• Convolutional Architecture Design
• Transfer Learning
• Object Detection Engineering
• Image Segmentation
• Edge Model Optimisation
• Vision MLOps

Why Attend This Course:

• Return with an Industrial Vision Model Evaluation Pack containing a trained model, evaluation evidence and a deployable package
• Stop guessing when training fails by reading loss curves, gradients and saliency maps to find the cause
• Move from image classification to detection, segmentation and tracking with code that can be reused on the next project
• Compare modelling choices with data scientists and engineers from manufacturing, energy, utilities and logistics during shared labs

Conclusion:

Deep learning delivers value in computer vision only when sound training practice, suitable architectures and honest evaluation meet a model that can run where the images are captured. Week one builds the mechanics: networks, optimisers, regularisation, convolutional backbones, transfer learning and sequence models. Week two adds what a short course leaves out: image pipelines, detection, segmentation, tracking, edge deployment, MLOps and responsible AI checks. The capstone applies all of it to an industrial dataset and produces an Industrial Vision Model Evaluation Pack.