Recuperação de informações de imagens através do uso de técnicas de visão computacional e extração de dados aplicadas a um conjunto de chaves de um claviculário real

Abstract

Through computer vision and fundamental approaches of the Knowledge Discovery in Databases (KDD) process, this work conducted digital image processing with the application of image acquisition, pre-processing, segmentation, feature extraction (mining), and evaluation steps to address challenges related to the identification of lost keys or possible security flaws in a real key cabinet. To achieve this, a dataset with labels and characteristics of various keys was created, which later served for information retrieval. The dataset consists of 350 examples of unique keys from a real key cabinet. Data collection was performed by taking photographs using a matrix with a solid and high-contrast background relative to the imaged object, along with uniform lighting to avoid unwanted shadows and enhance image definition. Subsequently, observations of the collected samples were made to identify relevant characteristics that could contribute to the proposed objective. With the goal of extracting features from the images in mind, the image segmentation phase was initiated, involving the selection of areas of interest, in this case, the area where the key’s secret (head) is located. Tools such as LabelImg were used for the selection of areas of interest, and Yolo was employed for training. After segmentation, image pre-processing was carried out by applying filters to the image. In this stage, libraries such as Rembg (based on neural networks) for background removal and OpenCV for converting the image to black and white (binary scale) were utilized. Feature extraction involved counting the pixels in the segmented region containing the key’s secret, resulting in a pixel intensity histogram. Finally, metrics such as chi-square, Manhattan, Euclidean, L∞, Canberra, and Mahalanobis were applied and compared with the aim of locating similar samples. As a result, in a set of 50 samples, an accuracy of 74% was achieved in identifying these keys. Regarding security, out of the 50 samples, 15 presented similar keys, totaling 20 instances. Among these, 4 were confirmed to be capable of opening the environments, corresponding to a 20% accuracy.