Data set is a most crucial aspect for the object recognition by the deep learning. The effect of training a deep learning model would be not good with an insufficient quantity of the data set. For example, there are two similar pictures which are captured from a video with a peacock and these two pictures are separated by one second in this video. Results show that the peacock was recognized by the model in the former picture but it failed in the latter picture due to the angle change of the peacock. In order to improve recognition effects of the model, we propose a system based on Gaussian Mixture Model and Speeded Up Robust Features to recognize peacocks in images. We also implement the prototype of this article and conduct a series of experiments to test the proposed solution. Furthermore, experimental results show the scheme did improve the accuracy of the complete training model.