In this paper, we propose a new framework hybridizing a Support Vector Machine (SVM), a Multi-Objective Genetic Algorithm (MOGA) and a Locality Sensitive Hashing (LSH). The goal is to tackle fine-grained classification challenges which means classifying many classes with high similarities between classes and poor similarities inside one class. SVM is used for its ability of learning multi-class problem from very few training data. MOGA is used for optimizing training samples used by SVM so as to improve its learning rate. As data define a discrete set of instances and not a continuous solution space, LSH is used for mapping "optimal solutions" obtained by MOGA onto the closest real instances contained in the dataset. We evaluate our method for content-based image classification on the standard image database Caltech256 (i.e. 30000 images distributed in 256 classes). Experiments shows that our method outperforms reference approaches.