Visual Question and Answering System for NLP and Computer Vision Problems

The purpose of this work is to develop a Visual question and answering system. Several neural networks of different architectures were traine. Ensembles of convolution neural networks and fully connected network, convolution neural networks and Long short-term memory, MobileNetv2 and Long short-term memory and MDETR are used. The MDETR model shown the best prediction accuracy. It made a lot of mistakes when it came to counting objects in a photo, but on other types of questions, it showed good results.