This article presents methods that allows to increase the accuracy of objects recognition systems and achieve better robustness of neural networks. The first proposed method of neural networks enhancement is the two-stage scheme of object recognition with two models trained on different sets of classes. The second method uses the new augmentation approach based on FGSM and JSMA algorithms to prepare the training dataset for model training. Combination of these methods allows us to achieve the ~ 20% increase of accuracy on complicated test images where false recognition occurred for original neural network model.