Weight-sharing multi-stage multi-scale ensemble convolutional neural network

Most of the existing convolutional neural networks (CNNs) ignore multi-scale features of input image to different extents. Thus they lack robustness to feature scale of the input image, which limits the generalization ability of the model. In addition, on the premise of large-scale data, in order to obtain higher image classification accuracy, CNNs generally require more layers and a huge amount of parameters, resulting in a higher cost of network training. To this end, a Weight-Sharing Multi-Stage Multi-Scale Ensemble Convolutional Neural Network (WSMSMSE-CNN) is proposed in this paper. The input image is pooled several times to obtain multi-scale images and sent to a multi-stage network. Each stage is a multi-layer multi-scale ensemble network consisting of Conv Block, Pooling layer and Dropout layer. Conv Blocks in the same stage are connected by pooling layers while those in different stage but at the same location share the same weights. In this way, multi-scale features of both the same image and scale features of multi-scale images are obtained. In addition, the large-sized convolutional kernel is replaced by a number of consecutive small-sized ones, which not only keep the receptive field unchanged, but also effectively control the number of parameters. Experimental results on CIFAR-10 and CIFAR-100 datasets verify that WSMSMSE-CNN not only has good robustness, but also requires fewer layers to obtain higher classification accuracy.