Design And Implementation Of Advertising Image Recognition System Based On Heterogeneous Computing

With the development of the Internet and the advent of the era of big data,the total amount of network data storage has exploded,human daily life has been inseparable with data,and everyone is a recipient of information.However,there are many disturbing advertisements in the massive network data,such as usury advertisements,scam advertisements including social accounts and advertisements of illegal websites abroad.At present,many websites use keyword screening to block advertisements in text format.Images are more concealed than text,so some advertisements adopt the method of embedding text in images,Therefore,many platforms can't distinguish whether the images uploaded by users to the server are advertisement pictures.When using chat software or social media,users will be disturbed by these advertisement pictures,which greatly affects their experience.In addition,the proliferation of advertising pictures may also provide a hotbed for illegal and criminal activities,such as gambling and cult activities.In order to solve the above problems,this paper designs and realizes an efficient and automated advertising image recognition system based on heterogeneous computing,which can automatically determine whether the picture uploaded by the user is an advertisement picture and give the recognition result.The mainly work is shown as the followings:First,the advertising image recognition system realizes the functions of human-machine interface,image text detection,image text recognition,and text classification.The system has a concise and easy-to-use human-machine interface,Users can select the system path of the data source to be identified,and the interface can update the recognition results in real time and feedback to users.Considering that the advertising image has a complex background,the image text detection function uses the natural scene text detection method.The system uses convolutional neural network to extract the features of the image to be identified,uses recurrent neural network BLSTM to learn the sequence characteristics of the proposals,and finally uses construction algorithm to get the text area.The implementation of image text recognition also uses convolutional neural networks and BLSTM,and uses CTC to align text.In order to avoid complex feature extraction steps,text classification uses a deep learning method to achieve end-to-end classification.Secondly,in order to solve the problems of slow image recognition and long CPU computing time,I use heterogeneous computing to accelerate the system.After analyzing the parallelism of the adopted algorithm and comparing the power consumption of the device,I select a heterogeneous computing system based on the OpenCL framework and use FPGA as the system's heterogeneous computing accelerated device.After testing,the system's detection rate of advertising image recognition reaches expectation.This system meets the requirements of system accuracy and recognition efficiency in demand analysis.Compared with the pure CPU solution,the heterogeneous computing solution improves the system efficiency.The system is of great significance to solve the problem of Internet advertising flooding.