My Skill Lab
Artificial IntelligenceIntermediate

Computer Vision Fundamentals

How machines see — from pixels to detection.

Taught by Asha Karki · AI Product Engineer

What you'll learn

  • Represent and manipulate images as NumPy arrays, including channels, dtypes, and pixel value ranges
  • Apply convolution filters by hand to extract edges, blur, and sharpen, and explain what each kernel does
  • Explain the core CNN building blocks — convolution, activation, pooling, and the feature hierarchy they learn
  • Build and train a small image classifier and read its accuracy, loss, and confusion matrix correctly
  • Run object detection to get bounding boxes, labels, and confidence scores, and interpret IoU and NMS
  • Use transfer learning and pretrained models (ResNet, YOLO) to ship a vision feature quickly and correctly

About this course

Computer vision turns raw pixels into decisions: what is in an image, where it is, and what to do about it. This course builds your intuition from the ground up — images as numeric arrays, the filters that extract edges and textures, why convolutional networks learn features automatically, and how to classify images, detect objects, and reuse powerful pretrained models without training from scratch. By the end you can load a model, run inference, and ship a working vision feature with realistic code.

Course content

8 lessons
  1. 1Images Are Just Arrays of NumbersPixels, channels, dtypes, and how an image becomes a tensor a computer can compute on.Preview
  2. 2Filters and Convolution by HandKernels, the sliding-window convolution operation, and how filters reveal edges, blur, and texture.Preview
  3. 3From Filters to Features: CNN IntuitionWhy CNNs learn their own filters, and how convolution, activation, and pooling stack into a feature hierarchy.
  4. 4Image Classification End to EndDatasets, train/validation split, the training loop, and reading accuracy, loss, and a confusion matrix.
  5. 5Object Detection: Boxes, Labels, and ConfidenceBeyond one label per image — predicting bounding boxes, scores, IoU, and non-max suppression.
  6. 6Transfer Learning: Stand on Pretrained ShouldersWhy you rarely train from scratch — feature extraction, fine-tuning, and when to use each.
  7. 7Preprocessing and Data Pipelines That Don't BiteResizing, normalization, channel order, and matching the exact transforms a pretrained model expects.
  8. 8Shipping a Vision Feature: Evaluation, Pitfalls, and ProductionHonest evaluation, common failure modes, latency and cost, and a checklist for putting vision in production.