Research And Implementation Of Brand LOGO Recognition Technology Based On Convolutional Neural Network

With the rise of e-commerce companies such as Alibaba,Amazon,JD,more and more users began to purchase goods on the e-commerce website through the pictures uploaded by the sellers,making the brand LOGO an important element of the electronic economy market.Although the research on image classification and recognition technology is relatively mature in technology,the identification of brand LOGO is still a very challenging and meaningful technology due to the diversity of usage scenarios,the identification of different data samples,and the limited memory resources on the deployed mobile devices.This thesis combines improved knowledge distillation algorithm with bias algorithm to solve the problem of deploying brand LOGO recognition system on mobile devices with limited memory resources.The thesis improves on the basis of the knowledge distillation algorithm and shares some parameter layers of the teacher network with the student network,so as to achieve the effect of simultaneous training of the teacher and student network.On the same reptile dataset,the average accuracy of the improved algorithm was increased from 0.89 to 0.91 compared to the knowledge distillation algorithm.This indicates that the student network has better recognition and expression ability based on network compression,saving training time cost for model iteration and update.In this thesis,the extracted features of the middle layer of the student network are further constrained,thereby realizing the feature migration from the middle layer of the teacher network to the middle layer of the student network.Therefore,the output of the student network is further approached to the output of the teacher network,so that the knowledge transfer from the teacher network to the student network can be better realized.By designing the bias vector and the loss function of the student network,this thesis can customize the accuracy of the target LOGO class without changing the original network scale.After the above algorithm is improved,the biased student network for the brand LOGO identification of the system is designed and implemented.