Image as numerical data (pixels, channels) in Computer Vision - Model Pipeline Trace

Model Pipeline - Image as numerical data (pixels, channels)

This pipeline shows how an image is turned into numbers that a computer can understand. It breaks down the image into pixels and color channels, then uses these numbers to train a model that learns to recognize patterns.

Data Flow - 4 Stages

1Input Image

1 image of 28x28 pixels with 3 color channels→Raw image loaded as height x width x channels→28 rows x 28 columns x 3 channels

A photo of a red apple represented as a 28x28 grid with RGB values

↓

2Normalization

28 rows x 28 columns x 3 channels→Scale pixel values from 0-255 to 0-1→28 rows x 28 columns x 3 channels

Pixel value 255 becomes 1.0, pixel value 0 stays 0.0

↓

3Flattening

28 rows x 28 columns x 3 channels→Convert 3D image data into 1D array for model input→2352 features (28*28*3)

The 3D pixel grid becomes a long list of 2352 numbers

↓

4Model Training

2352 features→Train a simple neural network to classify images→Output probabilities for each class

Model predicts 0.8 probability for 'apple', 0.2 for 'banana'

Training Trace - Epoch by Epoch

Epoch 1: ************ (loss=1.2)
Epoch 2: *********    (loss=0.9)
Epoch 3: *******      (loss=0.7)
Epoch 4: *****        (loss=0.5)
Epoch 5: ****         (loss=0.4)

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning, accuracy is low
2	0.9	0.6	Loss decreases, accuracy improves
3	0.7	0.72	Model is learning important features
4	0.5	0.82	Good improvement, model is converging
5	0.4	0.88	Loss low, accuracy high, training successful

Prediction Trace - 4 Layers

Layer 1: Input Image

Layer 2: Normalization

Layer 3: Flattening

Layer 4: Neural Network Prediction

Model Quiz - 3 Questions

What does normalization do to the image pixel values?

A Changes image size from 28x28 to 14x14

B Scales pixel values from 0-255 to 0-1

C Converts color image to black and white

D Flattens the image into a 1D array

Key Insight

Images are made of pixels arranged in height, width, and color channels. Turning these pixels into numbers lets a model learn patterns. Normalizing and flattening prepare the data so the model can understand it and make predictions.