Image Processing by Hand: Pixels, Convolution Kernels, Edges & Pooling

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• For ten colour pixels from a PixelForge test photo you calculate the perceived grey value with luma weights, compare it with a plain average, find the brightest pixel and work out how much memory a colour photo needs.
• Brighten and stretch a dim 8×8 photo of a lamp with clipping to 0–255, compare the mean brightness, count the original pixels per brightness bin with FREQUENCY and draw the histogram.
• Pick a global threshold for a photo of a coin on a dark table, turn the image into a 0/1 mask, measure how much of the image the coin covers and colour the mask.
• Slide a 3×3 Gaussian kernel over an 8×8 image with SUMPRODUCT, produce the 6×6 blurred output, compare it with a box blur and visualise the result with a colour scale.
• Sharpen one pixel with a sharpening kernel, then run the Sobel kernels over an image with a bright block to find vertical and horizontal edges and combine them into the gradient magnitude.
• For six convolution layers you calculate the output size and the number of activations, then compute a stride-2 convolution on a 5×5 image with OFFSET.
• Downsample an 8×8 feature map (a diagonal-edge detector's output) with 2×2 max pooling and average pooling, compute global average pooling and measure how much data pooling removes.
• Push a 6×6 handwritten "1" through a complete mini CNN: a vertical-line kernel with bias, ReLU, 2×2 max pooling and a dense layer with a sigmoid, and read the network's probability that the image shows a 1.