Nowadays, Depth Learning with Convolution Neural Networks (CNN) has become a well known method in object recognition task to generate descriptions and classify learned features and gradually took over. There are few CNN applied studies on leaves recognition and classifier task that mainly use the existing CNN architecture and pre-trained models. This paper proposes a CNN model for leaves classifier based on thresholding leaf pre-processing extract vein shape data and augmenting training data with reflection and rotation of the image. Preprocessing to reduce the storage capacity of data and increases the computational efficiency of the model. This model was experimented on collector leaves data set in the Mekong Delta of Vietnam, the Flavia leaf data set and the Swedish leaf data set. The classification results indicate that the proposed CNN model is effective for leaf recognition with an accuracy greater than 95%.