Similarity Measurement on Logo Image Using CBIR (Content Base Image Retrieval) and CNN ResNet-18 Architecture

Abstract

In this study aimed to measure the level of similarity between two logos, both those that look different and those that look the same. This can be realized by forming a logo image database that is stored in a logo image database derived from various existing logo image data. This research uses 4 logo images as data testing and 210 image data for the database as data training. All of the logo images used come from the Ministry of Law and Human Rights of the Republic of Indonesia (Kemenkumham RI) West Sumatra Regional Office. The size images are color images with a pixel size of 320 x 320 pixels, the purpose of which is for the process of dimensional uniformity of the images to be studied. This research uses Content Base Image Retrieval (CBIR) method to search for images from a large image database than using Convolutional Neural Network (CNN) type Residual Network (ResNet-18) Architecture to get the similarity score accurately. The result of this implementation is the formation of an automatic distribution of training images and validation images with 147 training image data values (70%) and 63 validation images (30%) of the 210 existing images. The result of this research is producing the algorithm to implement the method and the tool software application to measure the similarity of logo images. The accuracy of this tool is 93.65% with a total of 84 iterations.