Abstract: The convolution operators are the core building blocks of convolutional neural network, which enable the network to fuse the information of various layers of space and channels according to a certain perception field of view, and extract the characteristics of the information. However, adjacent pixels often have similar values in an image, which results in a large amount of redundant information in the output of the convolutional layer. In order to reduce redundant information and speed up model inference, many pooling layers are added to the convolutional neural network for reducing information dimensionality. Pooling has better dimensionality reduction effect on image features with the invariance of translation and rotation. And end-to-end model can be maintained compared with traditional dimensionality reduction methods. Therefore, a dimensionality reduction method is proposed based on the attention mechanism by using the pooling layer. In the process of feature extraction, the dimensionality reduction information from each layer’s are reused nonlinearly, so that the potential connections of information in different layers after dimensionality reduction can be learned. In order to obtain the characteristics of the input information, the proposed method focuses on the main texture of the target in the image, and then the low texture and background information of the target are combined. Based on the DLA-34 (deep layer aggregation) neural network, the dimensionality reduction method proposed in this paper and the others dimensionality reduction methods based on the maximum value and the average value are compared to deal with multiple sets on the CIFAR10 and CIFAR100 datasets, which proves the effectiveness of the new method.

Key words: deep learning, image classification, convolutional neural network, residual network, attentionmechanism