In diploma thesis we address the problem of counting objects in images. Manual object counting can be very time consuming and prone to human error when we have a large number of images with many objects. There are applications that offer counting different categories of objects with pre-trained models. We offer a slightly different solution. We propose the development of a web application that uses convolutional neural networks to count objects. With easy-to-use web application, we offer the user to train their own models and to use their own and pre-trained models to count objects in images. We build the web application using modern technologies such as React, NodeJS, Tornado and PyTorch. Finally, we want to know whether the built application really contributes to faster and more accurate counting and whether the application is well designed in terms of performance, accessibility, use of best practices and search engine optimization. We conclude that the application is well built, as we achieve excellent results in the testing. We also find that the application counts objects significantly faster than manually. We also compare the precision, recall and F1-score of the Faster R-CNN and FamNet methods. We find that the FamNet method has a higher recall value, while the Faster R-CNN method has higher precision and F1-score values.
|
