Deep neural networks have high classification capabilities however; a substantial disadvantage they hold is the necessity of a large classified database. Creating such databases requires a large amount of resources, and at times is impossible. The desire to harness the advantages of deep neural networks when there is a limited amount of classified examples has led to the development of different approaches to lowering the dependence of the classification ability in this factor. In this project, we focused on the approach of producing useful information for the wanted image database from the learning process of a deep convolutional neural network over a foreign image database. To improve the classification ability of the wanted dataset we examined the possibility of learning over two datasets, which are classified in different methods – a foreign image dataset classified properly and a foreign image dataset, which was classified randomly. We will display the process of choosing the networks architecture, building it and choosing its features. We will show that it is in-fact possible to achieve an improvement in the final classification by utilizing the knowledge acquired through the process of learning the features of a foreign image dataset.