Abstract:In popular lightweight convolutional neural networks (CNNs), pointwise convolution (PWC) layers for combining information occupy approximately 70% weights and computation, but depthwise convolution (DWC) layers for extracting spatial information only occupy less than 2% weights and computation. The weights and computation for extracting spatial information are not enough in lightweight CNNs. In this paper, we proposed expanding the number of channels and improving the extraction of spatial information by more efficient DWC instead of PWC. Firstly, the results of the proposed PSDNet demonstrate that DWC is more efficient than PWC for channel expansion and it can improve the accuracy of the network. Then, the efficient Depthwise-Pointwise-Depthwise (DPD) block is proposed by using DWC to expand channels. Different from the general bottleneck block, the DPD block consists of one PWC layer and two DWC layers. Four kinds of efficient lightweight DPDNets (DPDNet-G, DPDNet-A, DPDNet-C, DPDNet-D) are proposed by stacking different DPD blocks. To extract multi-scale features and achieve high accuracy, the pyramid DWC layer is used when channel expansion in DPDNet. Compared with common lightweight CNNs, DPDNets use more weights and computation in the DWC layer for extracting spatial information. Four competitive benchmark datasets (CIFAR-10, CIFAR-100, ImageNet, and PASCAL VOC) were used to verify the superiority of DPDNet. Experiments demonstrate that the proposed DPDNet has higher accuracy than MobileNet with a similar number of weights and computations. Furthermore, compared DPDNet with MobileNet, it can be found that improving the ratio of DWC to PWC can improve accuracy, which helps researchers to design better lightweight CNNs.

ConShuffleNet: an Efficient Convolutional Neural Network Based on ShuffleNetV2

ShuffleNeMt: modern lightweight convolutional neural network architecture

HENet:A Highly Efficient Convolutional Neural Networks Optimized for Accuracy, Speed and Storage

Research on Image Classification Method of Features of Combinatorial Convolution

LA-ShuffleNet: A Strong Convolutional Neural Network for Edge Computing Devices

ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices

ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

Dynamic Shuffle: An Efficient Channel Mixture Method

All You Need Is a Few Shifts: Designing Efficient Convolutional Neural Networks for Image Classification

Efficient channel expansion and pyramid depthwise-pointwise-depthwise neural networks

CondenseNet: An Efficient DenseNet Using Learned Group Convolutions

AddressNet: Shift-Based Primitives for Efficient Convolutional Neural Networks.

Generic Convolutional Neural Network with Random Pooling Area

Cascaded Subpatch Networks for Effective CNNs

Shift-based Primitives for Efficient Convolutional Neural Networks

Efficient densely connected convolutional neural networks

Deep Neural Network Acceleration with Sparse Prediction Layers

Hyperspectral Image Classification Based on a Shuffled Group Convolutional Neural Network with Transfer Learning

Go Wider: An Efficient Neural Network for Point Cloud Analysis via Group Convolutions

Depth-wise Decomposition for Accelerating Separable Convolutions in Efficient Convolutional Neural Networks