Distance and gender based smart advertising display system

The vast number of researchers have been focused on pattern recognition and computer vision fields in parallel with recent technological developments over the last two decades. Studies on these subjects have become widespread in recent years. In this thesis, a smart advertisement display system has been developed which feeds real time data from the camera source to get gender information and calculate distance from the camera source. The developed system has two main stages. Firstly, live broadcast stream, which gets data from the camera source, is handled frame by frame. Then, the face detection part is employed for predicting the gender and distance information. Secondly, detected face images along with the gender labels and distance values are sent to the advertisement display application via the web service and saved into the database. The advertisement system is run in a time counter and analyses the records from the database. The analysis process is based on statistical information such as gender label and distance value to determine advertisements having different levels of detail. Determination of gender information is very important for proper system operation. For this study, face detection and gender recognition classifiers were implemented. Fisherfaces, Support Vector Machines (SVM) and Convolutional Neural Networks (CNN) classifiers for gender recognition were trained. The SVM classifier with Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG) features were used at different times. Besides, various optimization works were performed by changing the parameters. One of the most popular deep learning methods, the CNN network type, was trained with GoogleNet architecture and the optimization was performed depending on the parameters. The LFW, IMDB and WIKI were used as training data sets and the FaceScrub was used as the test data set. Fisherfaces algorithm yielded an accuracy of 61.30%. When LBP feature extraction method is combined with SVM classifier, the accuracy rate of 75.32% was reached. The HOG feature extraction method with SVM was found to be more successful than LBP and reached an accuracy of 80.58%. Finally, CNN was determined to be the best classifier among all having an accuracy rate of 94.76%.