RELEVANCE FEEDBACK BY COMPOSITE FEEDBACK OBJECTS ON CBIR

• Main Author: Pardede, Jasman
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/53197
• Institution: Institut Teknologi Bandung

Content-based image retrieval (CBIR) systems trace, search, and retrieve images that exist in a digital image database, based on an analysis of the visual content of a given query image. The objectivity of the CBIR system is to find or discover images effectively from an image database based on the similarity of query image features provided by the user. Until now, the main problem of the CBIR system is the semantic gap. The semantic gap caused by two aspects, namely: the system perception aspect and the user perception aspect. The system perception aspect caused by the CBIR system describing the image based on the numerical value of the image pixels. In the beginning, the CBIR system designed following a computercentric approach that interprets image content based on low-level features extracted from image pixels. Humans can use high-level semantic concepts so that they can describe and mean the image content precisely. As a result, in the same image, there is a wide gap between computer interpretation and human interpretation. Various methods of feature extraction and similarity-measure developed to obtain the best representation of feature images. The feature extraction and the similarity-measure methods certainly cannot produce the best performance on all types of image databases. The semantic gap in the user perception aspect caused by the user sometimes does not have the initial idea about the image to be sought when searching for images. Search ideas come up based on the initial search results provided by the system. As a result, the image search is very subjective. By utilizing the outgrowth of deep learning technology to detect objects in images, the ability of the human brain to objects both recognize and detection, and the exploiting the advantages of Relevance Feedback (RF) techniques to improve queries, this research proposed an Object-Detection-Based CBIR (ODBCBIR) system. There are 2 (two) main contributions to this study, namely: (1) the proposed ODBCBIR system uses the equalset-based retrieval as a technique for measuring the similarity of the query image with each image in the image database and designing a storage structure for object feature extraction that can facilitate retrieval, (2) proposed object feedback on the ODBCBIR system uses to facilitate users to construct new queries based on object feature feedback using set iv operations. There are 5 (five) object feedback operations proposed in the ODBCBIR system, namely: AND, OR, XOR, DIF, and COS Feedback. The proposed ODBCBIR system was implemented in 3 (three) datasets, namely, the Wang, GHIM-10K, and, Corel-10K dataset. The proposed ODBCBIR system performance is measured based on the average value of accuracy, precision, recall, and F- Measure. The average accuracy values on the Wang, GHIM-10K, and, COREL-10K datasets were 0.997, 0.997, and 0.998, respectively. The average precision values on the Wang, GHIM-10K, and COREL-10K datasets was 0.983, 0.942, and 0.962, respectively. The average recall values on the Wang, GHIM-10K, and COREL-10K datasets was 0.955, 0.9480, and 0.888, respectively. The average F-Measure values on the Wang, GHIM-10K, and COREL-10K datasets was 0.983, 0.944, and 0.954, respectively. Based on the results of experiments conducted, it found that the proposed ODBCBIR system produced much better accuracy, precision, recall, and F-Measure performance compared to the state-of-the-art CBIR system. The technique of measuring object similarity with the equalset-based retrieval and the proposed object feature extraction storage structure can reduce the semantic gap from the aspect of system perception. The object feature feedback by using set operations in the proposed ODBCBIR system can construct similarities in perception between objects desired by the user and objects formulated by the system. Generated a new query can provide image retrieval results that match the user's perception of only 1 (one) time feedback. The feedback operators that are easy for respondents to understand are the AND and OR operators. Generated new queries by using set operations and based on the feedback of user object features can reduce the semantic gap in terms of user perception.