The Design And Implementation Of Deep Learning-based Picture Auto Tagging System

In this modern era of information explosion, people are taking a lot of digitalpictures all the time. And there are a large part of the pictures are taken from theuser’s smart mobile devices. In such a massive amount of data, in order to meetInternet user’s demand for finding and achiving pictures, keyword tagging forpictures becomes very necessary. The traditional method for tagging picture,often only use context-based text keywords to tag the picture, not the pictureitself. And the existing content-based image retrieval systems are not suitable forsolving the above problem: on the one hand, they are not providing tags, butproviding the similiar pictures; on the other hand, they are rely on network andserver computing resources, are not convenient on the mobile device.Aiming at the problem, we proposed a solution based on deep learningapproach. This solution can provide tagging suggest locally just after user takinga picture, the classify caculation are processed locally. It separates the servercaculation and the device caculation, thus easing the heavy computing on theservers caused by the large quantities of pictures. Unlike the previous machinelearning approaches, deep learning does not need to be assigned a method fordata representation, but will find out its own data representation method bygiving it a huge set of samples. This solution’s keypoint is to find a datarepresentation on PC platform, and use this representation on mobile platformfor classification computing locally.The main work of this paper includes:(1)Achieved a system based on deeplearning which separates the caculation of training and classifying to differentplatforms, connected by sharing a property configuration;(2)Implemented aconvolutional neural network trainer on PC platform, which written in Pythonand can be accelerated by using gpu computing;(3)Implemented a convolutionalneural network classifier on iOS platform, which written inobjective-c;(4)Developed a applicatoin on iOS platform, which can downloadclassification configurations, taking pictures and tag the pictures usingdownloaded configurations;(5)Used different images datasets to test and tunethe convolutional neural network’s training process, figured out a balanced,optimized neural network architecture. In this paper, based on software engineering development process, wecarried out a detailed analysis, design and implmentation explanation for thesystem. Currently the system is initially completed as a prototype system, whichhas achieved a series of full functionality from generate image sets to use theclassifier to tag a picture on mobile device.