基于多区域特征的深度卷积神经网络模型

Deep Convolutional Neural Network Model Based on Multi-Region Feature

摘要: 针对现有的深度卷积神经网络模型存在的网络结构复杂、计算量较大等问题，从而无法在实际中有较广泛的应用，该文提出一种基于多区域特征的深度卷积神经网络模型。模型首先对图像进行多区域划分，然后用标准卷积操作得到图像语义上下文信息，接着利用多区域的输入来学习上下文交互特征，通过把全局区域和多个分区域的空间信息级联再输入卷积层，以一种信息补充的方式提取图像的上下文特征信息，最后通过Softmax函数对图像进行分类。实验结果表明，该模型结构简单，参数量较少，且多区域特征融合上下文信息建模比单区域特征建模具有更好的鲁棒性和更高的分类精度。

Abstract: Aiming at the problems of complex network structure and large computational complexity of existing deep convolutional neural network models, it cannot be widely used in practice. This paper proposes a deep convolutional neural network model based on multi-region features. Firstly, the model divides the image into multiple regions and then uses the standard convolution operation to get the image semantic context information. Then, the multi-region input is used to learn the context interaction feature. By cascading and inputting the spatial information of the global region and the multiple sub-regions into the convolution layer, the context feature information of the image is extracted in an information supplement manner. Finally, the images are classified by the Softmax function. The experimental results show that the model has simple structure and few parameters, and the modeling with multi-region feature and context information fusion has better robustness and higher classification accuracy than that with single-region feature.