CV Workshop — Image Classifier

OpenCV + GoogLeNet (ImageNet) · Beginner Computer Vision Workshop

---

Getting Started

First, you will need to fork this repo, ensuring your fork is public

Then, work on the code in this repo, try to get as much done as you can!

Finally, answer the questions listed at the bottom of the README to be entered into the raffle.

And most importantly, have fun!

Setup

pip install opencv-python numpy

Then download the model files (one-time, ~50 MB):

python download_model.py

Sample images are already included in images/.

---

Repo Structure

cv-workshop/
├── utils.py                    ← Part 1: implement this first
├── model.py                    ← Part 2: implement this second
├── main.py                     ← Part 3: wire everything together
├── download_model.py           ← run once to fetch model files
├── synset_words.txt            ← 1000 ImageNet labels
├── deploy.prototxt             ← model architecture  (after download)
├── bvlc_googlenet.caffemodel   ← model weights       (after download)
├── images/
│   ├── dog.jpg
│   ├── cat.jpg
│   ├── car.jpg
│   └── bird.jpg

---

How to Work Through This

Work in order: utils.py → model.py → main.py

Each file has a self-test. Run it after you finish that file:

python utils.py    # must show all ✓ before moving to model.py
python model.py    # must show all ✓ before moving to main.py
python main.py     # runs the full pipeline

---

Running the Classifier

# Default image
python main.py

# Specify image and expected label
python main.py --image images/cat.jpg --label cat
python main.py --image images/car.jpg --label "sports car"

# Classify every image in a folder
python main.py --batch images/

---

The Pipeline

your_image.jpg
    ↓  load_image()
(H × W × 3)  BGR array
    ↓  preprocess()         grayscale → blur → Canny
(H × W)  binary edge map
    ↓  find_subject_contour()
largest qualifying contour
    ↓  crop_roi()
(h × w × 3)  color crop
    ↓  prepare_blob()
(1 × 3 × 224 × 224)  normalized tensor
    ↓  run_inference()
(1 × 1000)  confidence scores
    ↓  get_top_prediction()
"golden retriever"  94.3%
    ↓  draw_prediction()
annotated image on screen

---

Some ImageNet Categories to Try

Animals	Vehicles	Objects	Food
golden retriever	sports car	laptop	pizza
tabby cat	school bus	backpack	banana
bald eagle	ambulance	rocking chair	ice cream
hammerhead shark	mountain bike	sunglasses	coffee mug

Check synset_words.txt for the full list of 1000 valid labels.

---

Questions - MUST BE DONE TO ENTER RAFFLE

1. In your own words, explain why we preprocess the image with grayscale, blur, and edge detection before passing it to the model. What would happen if we skipped one of those steps?

2. When you ran your classifier on an image, what did it predict and how confident was it? Did the result surprise you — and if it got something wrong, why do you think that happened?

3. We focused on the top prediction (the supposed classification) — but the model outputs 1000 scores simultaneously. What does it mean that the scores for other classes are non-zero? What are those numbers telling you?

4. Where would you take this project next? Think about different models you could swap in, new kinds of images you'd want to classify, or features you'd add to make it more useful in the real world.

Reference Docs

• OpenCV DNN: https://docs.opencv.org/4.x/d6/d0f/group__dnn.html
• OpenCV all: https://docs.opencv.org/4.x/
• NumPy: https://numpy.org/doc/stable/reference/