IMAGE MATCHING MODELING USING THE GRAPH MATCHING APPROACH

• Main Author: Akmal
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/75248
• Institution: Institut Teknologi Bandung

Traditional pixel-based image analysis techniques do not extract information effectively because they only represent content. A very promising and challenging approach is to extract graphs from images which, in addition to showing the content, also show the relationships between the existing contents. Graph extraction is done through image segmentation process. The Minimum Spanning Tree superpixel segmentation method is one way to divide an image into regions. The regions obtained in the segmentation stage are represented as nodes and the relations between neighboring regions are represented as edges. This form of representation is called a Region Adjacency Graph (RAG). This graph representation and Graph Matching are used in image matching. The graph uses image feature representation by utilizing the region features and their interrelationships. It is assumed that the RAG obtained is a feature graph of the image, then a matching process is carried out from one graph as a test image with several image graphs in the image graph dataset. Graph Matching is used to calculate the similarity between image graphs. Image matching is used to find similar images by measuring how close the query image features are to other image features in the database based on the extracted features. Graph matching was performed using the Exact Graph Matching technique with the VF2 algorithm and inexact Graph Matching with the Graph Edit Distance (GED) algorithm. The main advantage of the Exact Graph Matching method is its strict definition and strong mathematical basis, while the inexact matching algorithm does not look for exactly the same correspondence between two graphs but looks for similarities between the two compared graphs. In this case, the cost (or distance) is calculated by calculating the difference in calculations between the corresponding attributes. The matcher will look for the mapping that minimizes the cost. This Graph Matching method will be applied in image retrieval. This is possible because image retrieval is a process of finding images in a database that are similar to the query image by measuring how close the feature values of the query image are to other images. Graph matching is used to calculate the similarity between image graphs. Image retrieval is currently dominated by approaches that combine several different representations or features. Besides the strengths of graph representation, there are weaknesses that occur in image processing using graph representation, so to improve the performance of the image matching process, basic low level features such as color, texture and shape are added. In this case the process being developed is the inexact Graph Matching method which will be combined with non-graph based methods in searching for visual image features. The Color Moment method is used to look for color features, the Gray Level Co-occurrence Matrix (GLCM) method to look for texture features and the Hu Moment method to look for shape features. The features obtained are compared between one image and another to obtain the similarity value of the two images using Euclidean Distance. Optimal weight of each feature is needed in combining image features. the. The weight of each feature in the combination of these features can be done manually or automatically. In this study, the Multi Layer Perceptron (MLP) Artificial Neural Network (MLP) method was used to obtain feature weights automatically and at the same time seek optimal weights. Testing of the developed method was applied to several data sets including artificial data sets, batik data sets, COIL-100 data sets and Wang data sets. The best Precision, Recall and F1_Score values were obtained for the 4 types of data sets as follows: for artificial data sets are 97.09, 88.27, 92.47 with CM + GLCM + HM method and threshold = 0.3, for batik data sets are 82.46, 82.41, 82.43 with the GLCM + HM method and threshold = 0.6, for the COIL-100 dataset are 90.26,53.99, 67.57 with the CM method and threshold = 0.3, and for the Wang dataset are 61.64, 40.52 and 48.90 with the CM + GLCM method and threshold = 0.5 . In general, the combination of the Graph Matching method with visual image features that can be weighted manually or automatically gets better performance results.