Image recognition assigns an image to one of several classes by using visual content. A typical example is labeling an image as “strawberry,” “plane,” or “dog.” section 1.31 section 1.27
Recognizing a cat
A cat can be identified through parts such as its tail, eyes, and whiskers, even when their exact positions or orientations differ. Image recognition therefore values the presence of useful parts more than their precise location. section 1.31
A convolutional neural network (CNN) receives an image as a table of pixel values. It applies learned filters that can blur regions, highlight colors, or detect sharp changes such as lines. Several convolutions produce a modified image, which is then downsampled to remove precise spatial detail. Repeating this process lets later filters detect increasingly abstract features built from earlier ones. section 1.31
Training depends on labeled examples and can fail in predictable ways:
- Supervised learning supplies many images paired with their correct categories; creating those labels can require manually tagging tens of thousands of images. section 1.27
- Generalization means working on new images rather than only memorizing training examples. section 1.27
- Overfitting memorizes training examples, while underfitting learns a process too simple to work effectively even on the training data. section 1.27
