Optimized CNN Based Image Recognition Through Target Region Selection

Image recognition has plateaued in the last few years. According to this research field, some complicated models typically combined feature extraction and classification models effectively. Moreover, many classic models have already achieved realistic recognition. However, there are still some drawbacks of traditional methods. On the one hand, some unrelated regions of learning instances are often used leading to ignorance of effective features. On the other hand, traditional CNN model don't consider the weights of learning instances which reduces the accuracy of image recognition. Aiming at the problems above, we proposed one optimized CNN based image recognition model. Firstly, target region selected by bottom-up region proposals contributes to retrieve the target region of each learning instance. Secondly, enhancement weight based model is used to optimize the CNN model contributing to make full use of different learning instances. At last, adequate experiments show our method's superiority, especially compared to some other traditional methods. (C) 2017 Published by Elsevier GmbH.