IMAGE SEARCH ENGINE WITH DIFFERENT IMAGE SIMILARITY SEMANTICS

• Main Author: RIFA’I - NIM : 13513073, WIWIT
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/31676
• Institution: Institut Teknologi Bandung

The increasing number of image collections makes image search engines increasingly needed to efficiently search for images in large image collections. Image collections that have descriptive text can utilize the descriptive text to be used in textbased search engines, but for text-free image collections it is necessary to use a content-based image search engine (CBIR) that is only based on visual information on the image. In CBIR, image similarity is based on features that can be extracted from images such as color, texture and shape. One of the main problems in CBIR is the semantic gap, the gap that occurs because of the limitations of extraction features in describing the expected semantics. Semantics that are used to compare the similarities of two images are very dependent on user’s perspective that tend to be subjective. Therefore, the similarity semantic of the image can vary greatly depending on the judgment and intent of the user. <br /> <br /> <br /> <br /> One approach in overcoming semantic gap is by weighting the image extraction feature. The weighting of these features determines the features that are considered more dominant in comparing image based on the intended image similarity semantic. Relevance feedback can be used to calculate the weighting of features based on feedback from users in order to get closer to the intended image similarity semantic. The relevance feedback method used in weighting calculations is Self Order Feature Reweighting. The Inverted Multi-Index is also used as an index data structure so that the image search process becomes more efficient. <br /> <br /> <br /> <br /> The test results show that the application of relevance feedback can provide increased accuracy but not too significant. The Inverted Multi-Index structure is able to increase the search speed for a small number of images retrieved. The weakness of the Inverted Multi-Index has the effect of decreasing accuracy and does not work well if the number of images taken is too much. <br />