Mixed attention dense network for sketch classification

Abstract

State-of-the-art convolutional neural networks (CNNs) on sketch classification cannot balance the expression ability of final feature vectors and the problems of gradient vanishing and network degradation. In order to improve the classification accuracy, we design a mixed attention dense network for sketch classification. According to the sparse characteristics of the sketch, this network uses overlapping pooling of a large size. In addition, dense blocks are added on the top of the middle convolutional layers to achieve feature reuse. Specifically, in order to extract more representative local, detail information, mixed attention is applied in the dense blocks. Finally, the center loss is combined with the softmax cross entropy loss to improve the classification accuracy. Through experiments, we compare our model with several state-of-the-art methods on the TU-Berlin dataset, and the experimental results demonstrate the effectiveness of our model.