【作者】 谷迪； 张东波； 王浩铭；

【Author】 Di Gu;Dongbo Zhang;Haomin Wang;College of Automation and Electronic Information, Xiangtan University;National Engineering Research Center of Robot Visual Perception and Control Technology, Hunan University;

【机构】 湘潭大学自动化与电子信息学院； 湖南大学机器人视觉感知与控制技术国家工程研究中心；

【摘要】 为了在资源受限条件下部署EfficientNet网络,本文提出对EfficientNet进行二值化改造的Efficient-BinaryNet模型。通过在骨干网络中反复堆叠包含DenseBlock和Improvement Block的模块来提升模型的特征图的容量和质量,同时引入逐点卷积跳跃连接防止信息丢失。此外,通过引入可学习参数,构建激活函数RSign和RPReLU,进一步增强二值权值的学习效果。在ImageNet1k数据集实验表明,在节省计算代价和提升精度方面,Efficient-BinaryN et优于目前很多先进的二值化网络模型。更多还原

【Abstract】 In order to deploy the EfficientNet network under resource-constrained conditions, we propos esthe Efficient-BinaryNet model, which is a binary transformation of EfficientNet. The capacity and quali ty of the feature representation of the model are enhanced by alternately the module including DenseBloc k and ImprovementBlock in the backbone network, while introducing point-wise convolution skip connecti ons to prevent information loss. In addition, by introducing learnable parameters and constructing activa tion functions RSign and RPReLU, the learning effect of binary weights is further enhanced. Experiment s on the ImageNet1k dataset show that Efficient-BinaryNet is superior to many current advanced binary n etwork models in terms of saving computational cost and improving accuracy.更多还原