Clothing Image Detection And Deployment Design Based On Deep Learning Framework

With the development of artificial intelligence and big data technology,people’s requirements for the efficiency of life and work continue to improve;and with the progress of science and technology,more and more businesses choose to sell clothes on the Internet,making more and more buyers begin to display their clothes correspondingly.As a result,massive clothing data is stored on the Internet.This also motivated the users to become more and more interested in fashion clothes,which makes fashion become a hot topic.However,in real life,fashion-related pictures and clothes are unavoidably deformed or concealed due to some objective factors,and the image styles of commercial clothing are also different.Understanding fashion images is still a big challenge in practical application.Therefore,the research of clothing image detection technology is of great significance to e-commerce platform and users.In recent years,the field of deep learning has made major progress,and it is more and more closely integrated with the fashion clothing.The deployment of model applications in real life has also become a hot research direction.While ensuring a certain accuracy of the model,it is also expected to continuously improve the running speed of the model and reduce the amount of network calculations.Based on this real-world demand,this paper focus on garment detection via deep learning framework.The main tasks are as follows:1.Under the framework of deep learning,the clothing image detection based on CenterNet is realized.The detection performance of the model is efficient as there is no need to set complex anchor-box,which can avoid many complexity and redundant calculations.A stacked hourglass network is also used so that a highly accurate detection model is realized;in addition,in order to reduce the parameters of the model,a residual network is used,which greatly improves the detection speed of the network.2.A new garment image detection network based on MobileNet V2 and CenterNet is proposed,and a feature pyramid network is added after feature extraction for feature fusion.In spite of losing a certain accuracy slightly,the detection speed of the model can be greatly accelerated,and the weight of model can also be reduced,which is convenient to complete the deployment on the platform with limited computing resources.3.In view of the low accuracy of the MobileNet version of CenterNet,a new clothing image detection network based on Vo VNet and CenterNet is proposed.According to the principle of hardware memory access cost,an efficient neural network,namely Vo VNet-slim-27,is used for feature extraction,which is relatively lightweight and good to make its calculation more efficient on the hardware platform as well as to improve both detection speed and accuracy.4.Two methods of network model deployment are proposed,which are based on Jetson nano to complete the portable implementation deployment of embedded hardware platform,as well as server deployment based on mobile application.In addition,universal technology stack uni-app and Java are selected so that the app development for mobile terminal and server development with rich computing resources are completed.