Future Technologies Conference (FTC) 2017 29-30 November 2017| Vancouver, Canada Image Retrieval using Graphs Daniel Valdes-Amaro * , Eduardo Lopez-Prieto * , Arturo Olvera-Lopez * and Carlos Guillen-Galvan † * Faculty of Computer Sciences Benem´ erita Universidad Aut´ onoma de Puebla Puebla, Mexico, 72570 Email: daniel.valdes@cs.buap.mx † Faculty of Mathematical Sciences Benem´ erita Universidad Aut´ onoma de Puebla Puebla, Mexico, 72570 Email: carlosguillen.galvan@gmail.com Abstract—Content-Based Image Retrieval has been a fast advancing area since the 1990s decade, with the Internet growth and the technology available that provides an easy access to acquire images. Hence, image data requires to be organized so that image database queries can result as fast as possible, even though the many possible topics are available now a days. In this paper, we introduce a novel technique based on global image features by interest point detection and using graph theory, particularly Delaunay triangulations to obtain a graph that can be measured for comparison. The technique shows promising results and can be regarded as flexible in the sense that parts can be adapted or upgraded to achieve better performance. Keywords—Content-based image retrieval; interest point detec- tion algorithms; graph theory; delaunay triangulations I. I NTRODUCTION Content-Based Image Retrieval (CBIR) refers to techniques of automatic retrieval of images from a database based on its contents, namely, visual features that describe the image contents [1]. In CBIR two different main approaches can be considered, discreet and continuous. The first one is inspired by textual information retrieval and uses such metrics for retrieval. On the other hand the second approach represents the images as feature vector, so these features are compared using different distance measures [2]. As a result, the use of visual features is a key component for the representation. Common features include colour, shape and texture and are called low level features, meanwhile high level features recall for a combination of low level features and a predefined model [3]. An open question here is then, which features are more suitable for image retrieval tasks. Part of the answer can be determined by the use of global or local image description features, that rely on the objects in images consisting of parts that can be modelled or not with varying degrees of independence [4]. Here, we introduce a novel technique for Content-Based Image Retrieval that works with the image in a global fashion using interest point detection algorithms and graph theory, in particular, Delaunay triangulations and graph measurements. Fig. 1 summarises the procedure. Some of the previous works similar to our approach (either in the use of interest points or graph theory) include the work in [5], where two methods to find interest points and a Gabor filter are used to create descriptions based on textures for image indexation that generate good results according to the authors over different image data bases. In [6], the image is divided in a series of subregions with a different area according to the distribution the interest points so local features are obtained, so using multiple instances the method improves the image retrieval precision. In [7] a scientific- cultural collection of hebraic tombstones is used to evaluate the performance of different interest point algorithms (SIFT- SIFT, SIFT-BRISK, SURF- SIFT, SURF-SURF, SURF-BRISK and CenSurE-SIFT) to implement CBIR techniques. These systems seek to classify images the way the human being does, but in a different fashion, in [8] a method is described were they convert of a given video or multidimensional imaging (MDI) into time series, followed by the conversion of it into a network, and finally analysing the network metrics they are able to identify specific topological metrics which can be regarded as discriminators for different inputs. The remainder of this paper is organized as follows. Section 2 gives details of the proposed method, in Section 3 the experiments are shown, and finally, in Section 4 some conclusions are given along with future work extensions. II. PROPOSED METHOD The first step is to select a query image, then the dominant RGB channel of the image is obtained in order to create a first filter for the images, remaining the ones from the data base with the same dominant colour as the query (Fig. 1(a)). After that, we proceed to extract image features, in this case and as mention before, global image features are considered. The concept of feature descriptors refers to methods that aim to calculate abstractions of information in an image, in addition to making decisions at each point in the image if there is a characteristic of a given type at that time [9]. In this case image interest points detection algorithms are applied in order to obtain a set of meaningful points (Fig. 1(b)). Interest points can be regarded as points that have a well-defined position in image space, stable under local and global perturbations and include an attribute of scale, to make it possible to compute interest points from images at different scales [10]. Five different descriptors were used in this work, in order to answer which feature is better for this task. First we have the Brisk feature detector, which relates key points in cases without sufficient a priori knowledge of the scene and angle of the camera. Fast, which is a corner 1022 | Page