Um sistema para organização de acervos baseado em visão computacional

Abstract

Search engines have become essential with the massive increase in the amount of information generated everyday in the internet. According to Cisco (2019) 82% of these information consists of images and videos. Therefore the development of robusts mechanisms to index this type of information, extract characteristics and search this visual information is necessary. Individually, each person has their personal needs for producing images and videos, depending on their lifestyle, but many have a smartphone in hand. For people with busy lives, who have memory problems or many things to remember, having a tool quickly available can be very useful in remembering things, like the date of a meeting or where you left your car keys. The objectives of this work are the development and implementation of a prototype of an application for the indoor organization of personal collections using computer vision and auxiliary tools for text and voice processing. The application was designed to be an easy and practical tool for using to organize personal belongings, whether or not consisting of a collection. In a literature search for a starting point to the design of this prototype, several methods and tools that could assist in its feasibility were found, however, no work with a similar proposal was identified. Therefore, many of the methods in this work had to be investigated, implemented and tested during the development of the prototype. The methods selected for the composition of the prototype are: ORB feature extractor, the vocabulary tree structure for storing descriptors and Google Speech-To-Text for speech processing. The validation was carried out with a large number of combinations of test cases so that one could have an idea of the capabilities and imperfections of the tool that was developed. Among them, what stands out is the ability to accept several consecutive insertions without the need for retraining at each insertion, however needing a regrouping of the tree descriptors, so that the processing time does not increase and the level of quality achieved , of more than 80 %, remain in an environment with an immense number of variables, such as: camera focus, motion blur, lighting, shadow, etc. As a conclusion of this work, which, although it still has many points for improvement, we have that the proposed objective was fulfilled and that, given the required improvements, the prototype developed could be turned into a real product capable of generating a positive impact on the life of the target audience.