[en] The work presented in this thesis is motivated by the problem of automatic image classification. Image classification methods seek to automatically classify previously unseen images using databases of labeled images provided by human experts. The main contribution of this thesis is a novel approach for image classification that has been shown to perform well on a variety of tasks. It uses some recent machine learning algorithms based on ensembles of decision trees that we applied directly on pixel values. We combine it with techniques of random extraction and transformation of subwindows from images so as to improve robustness to certain image transformations. The method has been evaluated on 7 publicly available datasets corresponding to various image classification tasks: recognition of handwritten digits, faces, 3D objects, textures, buildings, themes, or landscapes. Some of these datasets contain images representing widely varying conditions: occlusions, cluttered background, illumination, viewpoint, orientation, and scale changes. The accuracy of our method is generally comparable with the state of the art and it is particularly attractive in terms of computational efficiency.