DenseNet模型轻量化改进研究

Lightweight improvement research of DenseNet model

针对深层DenseNet模型在小型数据集上的过拟合问题,提出了一种改进的轻量化DenseNet模型.首先,优化网络中密集连接块(Dense Block)数量和其内部网络结构;然后,提出一种自适应池化层方法,解决改进网络的特征图分辨率适应问题;最后,加入Skip Layer模块增强密集连接块间特征信息流通.实验结果表明,改进方法能够减少模型的参数量和计算量,有效解决了深层DenseNet的过拟合问题.

Aiming at the overfitting problem of deep DenseNet model on small-scale data sets,an improved lightweight DenseNet model is proposed in this paper.Firstly,we optimized the number of DenseBlock and its internal network structure.Then,an adaptive pooling layer method is proposed to solve the problem of adapting the resolution of the feature map of the new network.Finally,we added the SkipLayer to enhance the flow of feature information between DenseBlocks.The experimental results illustrated that the new method can reduced the parameter and calculation amount of the model,and solve the over-fitting problem of deep DenseNet effectively.