传统神经网络在提取图片特征时,存在网络参数量过多、计算复杂度过高等问题,且网络对移动设备的存储空间和计算能力有着极高的要求,限制了网络图像分类方法在嵌入式设备上的发展与应用.为此,提出一种基于分组神经网络卷积与深度可分离卷积的轻量级图像分类方法.引入分组参数g和扩展参数n,实现网络精度与复杂度之间的平衡,并将分组卷积中普通卷积替换成深度可分离卷积,有效减少了神经网络中的参数量.实验结果表明,与MobileNet V1 方法相比,所提方法在数据集CIFAR-10上的准确率提高了4.75%,达到88.66%,且参数量仅为1.5 M.
An Image Classification Method Based on Multi-branch Lightweight Neural Networks
Traditional neural networks have problems such as excessive network parameters and high computa-tional complexity when extracting image features.Additionally,networks have extremely high requirements for stor-age space and computing power mobile devices,which limits the development and application of image classification on embedded devices.Therefore,a lightweight image classification method based on grouped neural network convo-lution and depthwise separable convolution was proposed.The proposed method introduced grouping parameter g and extension parameter n to achieve a balance between network accuracy and complexity,and replaced ordinary convolutions in group convolutions with depthwise separable convolutions,effectively reducing the number of param-eters in neural networks.The experimental results show that compared with MobileNet V1,the proposed method in-creases accuracy by 4.75%on the dataset CIFAR10,reaching 88.66%,with only 1.5 M of parameters.
deep learninglightweight neural networkgroup convolutiondepthwise separable convolutionimage classification
NETL
NSTL
万方数据
