Recently, image classification draw attentions of many researchers. The need of object recognition grows drastically, especially in the context of biometric, biomedical imaging and real time scene understanding. Computer vision task is the most challenging in machine learning. For that reason, it's fundamental to tackle this concern using appropriate clustering and classification techniques. However, the quest for the best unsupervised features extraction remain an open problem even if CNNs reach a remarkable success, establishing new state-of-the-art. In this context, we study from an acute insight standpoint the standard clustering models K-means, GMM and Naive Bayes classification algorithm in order to draw conclusion and underline their limits for such complicated tasks. To what extent are k-means and GMM efficient ? Why they fail and how to circumvent their weaknesses.