CNN Image Classiffcation on Military Pictures

Abstract

This project is part of a pipeline under the title "Adopt a bullet" that aims at gathering information of di erent weapon transports by using AItechniques. One stage in this pipeline consists of identifying the information that is relevant for solving this problem. To approach this stage I implemented a convolutional neural network (CNN) and trained it on a large set of images. The research question was, if it would be able to distinguish between images depicting military armoury and those that are not reliably. In this case, images of tanks have been used for training. After an initial training over 10 epochs, an accuracy of 74.33% was achieved. A second, smaller CNN was trained in an attempt to prevent over tting. This second CNN achieved a nal accuracy of 82.05%. This is a good result, but over tting still occurred. Further experimentation on its prevention as well as further eld-testing of the CNN is recommended, for example by applying it to a web-crawler.