| <— 6_1_Image_denoising.md | Зміст | 6_3_Object_detection.md —> |
6.2 Image classification
Image classification is the simplest strategy for extracting semantics from an image and consists of predicting a class from a finite, predefined number of classes, given an input image.
The standard models for this task are convolutional networks, such as ResNets (see § 5.2), and attention-based models such as ViT (see § 5.3). Those models generate a vector of logits with as many dimensions as there are classes.
The training procedure simply minimizes the cross-entropy loss (see § 3.1). Usually, performance can be improved with data augmentation, which consists of modifying the training samples with hand-designed random transformations that do not change the semantic content of the image, such as cropping, scaling, mirroring, or color changes.