A convolution slides a small grid of weights (a kernel) over an image. At each position it multiplies the pixels under the window by the weights and adds them up, giving one number of the output, the feature map. The map is large where the image looks like the kernel’s pattern, such as an edge. ReLU then keeps only the positive responses, and max-pooling shrinks the map by keeping the strongest response in each 2×2 block.