A neural network structure with wide range scale robustness

In the past few years, convolutional neural network-ks(CNN) has made great progress in various computer vision tasks, but its ability to tolerate scale variations is limited. For solving this problem, a common solution is making the model bigger first, and then trains it with data augmentation using extensive scale-jittering. This method greatly increased the study requirement. In this paper, we propose a multi-column structure of CNN, and experiment it at a basic neural network. The structure can effectively solve the problem of scale robustness in target recognition, and almost haven't any increase in study requirement. Especially, our structure is particularly effective when dealing with a wide range of scale-variant problem.