Addestramento di reti neurali su dati sintetici per rilevamento prodotti nella distribuzione al dettaglio ​

Artificial neural networks, or simply neural networks, are a machine learning algorithm that in the last decade found its use in almost all of our life fields. These algorithms have been going in and out of fashion for more than 70 years but thanks to modern advancements in hardware, e.g. GPUs, and the large amount of data we generate every day, neural networks have seen their resurgence at the beginning of the 21st century. Not all this data is in fact useable, since the major achievements of deep learning have occurred in the supervised field and it requires labelled data. Labelling real data is a heavily time-consuming task and not always error-free. For these reasons generating synthetic datasets resembling real ones became a very popular choice where a lot of research is currently done. In the consumer retail domain, operators have to fill reports about products displacements on the shelves, their prices, etc. This task is currently done by hand which requires a lot of time. In this project we want to simplify this work through its automation, which we achieved using an artificial neural network that will be used to recognize products in images made to the shelves and pre-fill the reports with the output generated data. To train this network, we needed labelled data, however there is not a publicly available dataset that is labelled for our high level of detail (EAN) work, for this reason we had to create one ourself. Said this, our two main goals are: automation of the production of synthetic labelled data and the training of a network to achieve object-detection in the retail domain. This thesis aims at automating the shelf reporting by means of a deep neural based approach. To solve the problem of generating synthetic labelled data we used Blender with which we created 3D models of the objects we want to detect. Putting these objects together in a scene allows the creation of renders that will be used as training data. To perform the object-detection task we used a YOLO network. The main reason is that we need a fast network since it will be used in mobile devices. The usage of AI in the mobile field is known as Edge AI, and this solution is becoming more popular each day since it provides multiple advantages over the usage of a central server. By performing the computation on our devices, no connection is required to the server, this reduces the latency and the overall cost. However, smartphones also have their limits we need to keep an eye on, such as battery life, low memory and computing power. Experimental results show that the proposed neural network achieve on average a recall score of 70-75% with more than 65% of precision taking an average of 70ms to perform a detection of an image on a system powered by a Tesla T4.