A filter that slides and detects
A convolutional filter, or kernel, is a small grid of numbers, often just 3 by 3, that slides across the image and, at each position, multiplies its values against the pixels underneath and sums the result. A filter tuned to detect vertical edges produces a high output value wherever the image has a sharp left-to-right brightness change, and near zero elsewhere. Stack many different filters, some sensitive to horizontal edges, some to color transitions, some to specific textures, and you get a set of feature maps that highlight different low-level patterns across the image.
This is the core operation inside a convolutional neural network, CNN, an architecture like ResNet or the earlier AlexNet that made deep learning dominant in vision starting around 2012. Critically, the filter values aren't hand-designed, they're learned automatically during training, adjusted through backpropagation until they detect whatever patterns actually help the network make correct predictions.