This chapter deals with object recognition in images involving a weakly supervised classification model. In weakly supervised learning, the label information of the training dataset is provided as a prior knowledge for each class. This prior knowledge is coming from a global proportion annotation of images. In this chapter, we compare three opposed classification models in a weakly supervised classification issue: a generative model, a discriminative model and a model based on random forests. Models are first introduced and discussed, and an application to fisheries acoustics is presented. Experiments show that random forests outperform discriminative and generative models in supervised learning but random forests are not robust to high complexity class proportions. Finally, a compromise is achieved by taking a combination of classifiers that keeps the accuracy of random forests and exploits the robustness of discriminative models.