Abstract: Aiming at the problem that the traditional deep convolutional neural network has poor classification accuracy and large amount of parameters, which is difficult to deploy in memory-constrained devices, a multi-scale parallel fusion lightweight convolutional neural network architecture PL-Net is proposed. Firstly, the upper output feature map is sent to two different scales of the depth separable convolution layer, and then the parallel output is cross-fused with the feature information, and with the residual learning, a parallel lightweight module PL-Module is designed. To better extract the feature information, the scale-dimensional reduction convolutional module(SR-Module) is proposed to replace the traditional pooling layer. Finally, the above two modules are stacked on each other to construct a lightweight network. In the experimental phase, training and testing are performed on the CIFAR10,Caltech 256 and 101_food data sets. The results show that compared with the traditional CNN,MobileNet-V2 and Squeezenet networks of the same scale, PL-Net improves the classification accuracy of the network while reducing the amount of network parameters, and is suitable for deployment on memory-constrained devices.

Key words: convolutional neural network, depthwise separable convolutions, residual learning, parallel convolution