Improvement And Implementation Of Lightweight Neural Network For Mobile Devices

With the rapid development of convolutional neural networks and mobile internet devices in recent years,there is an increasing demand for efficient lightweight convolutional networks that can be applied to mobile devices.This paper mainly proposes a new type of lightweight convolutional neural network-LDNet.Based on the traditional linear and light convolutional neural networks,LDNet optimizes the network by improving the connection mode of the backbone network and increasing the processing mechanism between modules.Compared with traditional convolutional neural networks,LDNet has the advantages of light weight,real-time and low overhead.LDNet is based on the traditional lightweight convolutional neural network structure,and has made corresponding structural adjustments to the local modules and backbone network parts,highlighting the effective information of the feature maps and obtaining higher quality image semantics.The main improvements are: In each module of the basic backbone network,attention mechanism module is added to achieve the purpose of enhancing important information,weakening unnecessary information and filtering noisy information.The attention mechanism module consists of channel attention and spatial attention components,which improves the quality of the feature map.(1)It is used to enhance the main feature map semantic information by adding a attention module to each module in the basic backbone network,weaken the secondary semantic information and shield the noise.The attention mechanismmodule starts from the two dimensions:channel attention and space attention,and makes the semantic quality of the obtained feature map is greatly improved.(2)LDNet's backbone network uses a new dense connectivity structure.Through the new hybrid connection,the feature semantics of each layer of LDNet's backbone network can make up for each other,and finally obtain a high-quality feature map with richer semantics.Through the above two major improvements and other local optimizations,the LDNet network significantly improves the recognition accuracy of the model with a small increase in parameters.Under the framework of TensorFlow,we tested the LDNet.The experimental results show that the LDNet with the dense connection of the attention mechanism and the backbone network is more mainstream than the MobileNet and ShuffleNet under the premise ? The data set is the Cifar-10 and Cifar-100 data sets.The lightweight neural network operates more efficiently and performs better in terms of recognition accuracy.