Within the field of pattern analysis, the Fisher kernel is a powerful framework which combines the strengths of generative and discriminative approaches. The idea is to represent a signal with a gradient vector derived from a generative probability model. We used this framework to describe the content of images. In our case, the input signal is one or multiple low-level local feature vectors extracted from the image and the underlying generative model is a Gaussian mixture model which approximates the distribution of low-level features in any image. We will explain how we successfully applied this idea to several image analysis problems including image categorization (i.e. annotation), image semantic segmentation and image thumbnailing.