A method of training convolutional neural network with only less positive samples

In this article an new approach is proposed to train a convolutional neural network with only very less training samples. In order to recognize the objects represented by the few positive samples in high accuracy, a set of rules are used to automatically extend the less training samples to a large number of training samples which cover the sample space as complete as possible. The methods of sample generation include shifting objects in the sample pictures with a slight up, down, left and right, expanding and thinning target objects, magnifying and shrinking the target objects and so on. A large number of effective positive samples are obtained using the methods above. A large number of negative samples are also generated for a CNN network model. Finally the trained network achieves excellent performance in terms of recognition rates and recognition reliability for objects represented by the less positive samples. Therefore, using the proposed approach the CNN can not only be trained to recognize digital numbers and it can also be trained for judgment of print quality of digital numbers.