Research And Application Of Light-weight-based Convolutional Neural Network On Sports Image Classification

With the vigorous development of the sports industry,the amount of sports image data is increasing exponentially.It is very important to effectively classify sports images,which can not only facilitate users to quickly retrieve and access sports images,but also help the staff with the data storage and management of sports images.And it contributes to the intelligent development of the sports industry.At present,many convolutional neural networks have achieved very good accuracy in image classification tasks,but the size of the network model and the amount of calculation have also increased,which requires computer equipment to have powerful computing power and memory,which is limited to a certain extent Therefore,convolutional neural networks are widely used on computer equipment with limited resources.In order to apply sports image classification to mobile devices with limited performance and resources,this paper makes light-weight research and improvement on convolutional neural networks to obtain a more streamlined network model and deploy it on Android devices for application.The main work of the thesis includes the following aspects:(1)Construct a multi-category sports image data set.Preliminary screening of the collected candidate pictures,eliminating some non-conforming and low-quality pictures,using bilinear interpolation algorithm to normalize the size of the processed image,and then use the data enhancement method to expand the data.Solve the problem of insufficient samples,and finally get a relatively reasonable sports image data set.(2)Use shallow ideas to simplify the network model,and the size of the convolution kernel is mainly 3×3.In order to improve the classification accuracy of the network model,a large number of comparative experiments are used to determine the activation function,gradient descent algorithm,learning rate,dropout discard rate and batch size.Finally,several experiments were performed on the shallow network model in different proportions of the training set and validation set.The experimental results verify the stability of the model,and the comparison experiments with different networks further verify the speed and accuracy of the network model have been improved to a certain extent.(3)Light weight improvements are made based on the structure of the shallow network model.We proposed a module which embeds depth separable convolution in the group convolution.One side of the group uses depth separable convolution with convolution kernels is set as 5×5,and the other side of group uses two 3×3 convolutions.Using the convolution kernels of different sizes to obtain richer image information.And then using channel shuffling for feature fusion and combines 1×1 convolution to reduce the dimensionality.Through comparative experiments with common lightweight-based convolutional neural network,it is verified that the network is more efficient on low-performance devices.The improved light-weight-based convolutional neural network has achieved higher classification accuracy while increasing the speed.Finally,the network is deployed on the mobile terminal for practical application and the category judgment results are displayed,further verifying the efficiency of the improved lightweight network model.