Object classification in images

A computer implemented method of classifying an object in an image comprises: identifying, using machine-learned local feature identification models 106, a plurality of local features of an object 108, each local feature comprising a feature identity (feature class) and area (e.g. bounding box); comparing the local features to a plurality of sets of masked images of candidate objects 110, wherein each set of masked images corresponds to a respective candidate object and comprises a plurality of viewing angles and each masked image comprises a plurality of feasible search zones for local features of the candidate object; and where the comparison is based on an overlap between the identified local features and feasible search zones; and determining a classification of the object 112 based on the comparison. The models comprise region based convolutional neural networks trained to detect local features. When identifying vehicles these features may comprise wheels; windscreens; windows/viewing slits; exhausts; tracks; weapon systems; turrets; engine grills; antennae; or the like. The technique reduces the effect of attack vectors such as adversarial noise and imagery, improves handling of high object variance and partial occlusion and facilitates low shot learning without the need for costly retraining.