A Novel Similarity Measure For Trademark Images Retrieval Based On Genetic Algorithm

Measuring perceptual similarity and defining an appropriate similarity measure between trademark images remain largely unanswered. Most researchers use the Euclidean distance or direct Hausdorff distance.Euclidean distance between any two p-dimensional patterns is considered as the difference between those two features in magnitude, rather than just the correlation of the features. Hence this measure has two main drawbacks when using Euclidean and direct Hausdorff metrics. Firstly, the largest-scaled feature tends to dominate the others and secondly, treating all features in the same way could affect the retrieval of results as images are transformed in different ways. Regarding, the direct Hausdorff distance, this distance plays an important role in image matching. However, we have to deal with image matching problems that occur when we use this measure such as image noise, occlusion and clutter. On the other hand the computation involved for the direct Hausdorff distance is time-consuming as you need to search for each pixel in the set of I_q, the closest of the set of I_s. where I_q is the image query and I_s is image similar.For a good content-based image retrieval system, we have to take into account three things, we must first choose a good similarity measurement, secondly, know the relationship between images features and their similarity measurements and finally have a good method for features weighting assignment. We propose the use of normalized cosine and Euclidean distance as finite set points instead of a set of image pixels in our shape based retrieval system, used in the application of trademark image retrieval.The proposed measure takes into account the integration of global features (invariants moment and eccentricity) and local features (entropy histogram and distance histogram), unlike other features extraction techniques. We have discovered two major problems. Firstly, some researchers often focus on using a single feature, e.g., Fourier descriptors, invariant moments or Zernike moments, without combining them for possible better results. Secondly, even if they combine the shape features, the weighting factors assigned to the various shape features are often determined in a static way.In order to narrow the search for trademark images and to reduce the time for searching for these images the database holding them is indexed. The selected images must have entropy value ranked in [0, 0.5] or [0.5, 1]. After that, the similarity function is computed between images from the database subset and the I_q, We have used the retrieval efficiency equation in order to test the accuracy of our method. Our thesis research carried out a novel similarity measure using a normalized Hausdorff distance for trademark retrieval that focuses on accuracy, robustness and execution cost in distance similarity computation between two trademark images. Our similarity measure approach uses two algorithms the sub-block-based trademark image retrieval algorithm under the polar coordinate system and Genetic algorithm. After normalizing the trademark image, the minimum circumscribed circle of the object is firstly set as the object region; then the object region is divided into some blocks under the polar coordinate, and the relevant shape histograms is computed and smoothed. We use learning method for finding the weighting factors in the dissimilarity function through retrieval of a large set of trademark images. A genetic algorithm is applied to decide the best weight factors distribution.From our simulation results, our aim of retrieving all best trademarks images is achieved because the percentage of the non-retrieved images in the top 20 matches is null for normalized Hausdorff distance. Hence, this is the key of the performance of normalized Hausdorff distance technique to the normalization of cosine and Euclidian distances technique. For cropped K = 10, it looses 4%, when K = 15 it looses 8%. Finally when K = 20, where K is the size of the short list of the best-first images retrieved, the direct Hausdorff distance looses a significant percentage. From the experimental results, the normalized Hausdorff distance by taking the minimum of the normalized Euclidean and cosine distances, outperformed the direct Hausdorff distance by decreasing the percentage of missing images from 20% down to 2%).The obtained results showed that normalized Hausdorff distance of cosine and Euclidean distances provide a significant improvement in retrieval accuracy and is robust against shape invariant transformations. Moreover, all the target images were ranked in the top twenty positions. This reveals that the dissimilarity function satisfied the human perception quite well.