بازیابی شباهت مبتنی بر محتوای علائم تجاری با استفاده از بازخورد مربوطه

This paper addresses the problem of how to efficiently and effectively retrieve images similar to a query from a trademark database purely on the basis of low-level feature analysis. It investigates the hypothesis that the low-level image features used to index the trademark images can be correlated with image contents by applying a relevance feedback mechanism that evaluates the feature distributions of the images the user has judged relevant, or not relevant and dynamically updates both the similarity measure and query in order to better represent the user's particular information needs. Experimental results on a database of 1100 trademarks are reported and commented.

The content-based retrieval of trademarks is “extremely challenging and instructive to study”, due to the high complexity and diversity of the data involved, also often composed of several distinct components [1]. Our study has addressed the problem of how to efficiently and effectively retrieve images similar to a query from a trademark database purely on the basis of low-level feature analysis. As already pointed out by several authors [1] and [2], perceptually similar images are not necessarily similar in terms of low-level features. We have investigated the hypothesis that the low-level features used to index the images can be correlated with their semantic contents by applying a relevance feedback mechanism. A few applications designed specifically for the registration of trademarks are available. Wu et al. have developed a prototype system, STAR, using their content-based retrieval engine for multimedia information systems [2], [3] and [4]. Eakins et al. have developed a prototype system (ARTISAN) for the UK Patent Office Trade Marks Registry, to retrieve trademarks when these consist of abstract geometric designs [5]. Another system, called TRADEMARK, operates on the trademark database of the Patent Office of Japan [6]. A detailed analysis of the problems involved in trademark registration can be found in a recent paper by Jain and Vailaya [1]. These authors propose a computational strategy in which multiple feature description schemes of the same visual cue (shape) are used to improve retrieval accuracy without significantly increasing computational costs. At the first stage of processing (pruning) Jain and Vailaya represent the trademark images in terms of invariant moments and the histogram of the edge directions, integrating the dissimilarity of these features by a weighted mean. A small set of plausible candidates is then presented to a detail matcher based on deformable templates to eliminate false matches. This second phase makes it possible to eliminate the false matches, but cannot cope with trademarks that have not been retrieved in the first stage, although actually perceptually similar to the query. We have attempted to improve the effectiveness of the first stage of retrieval by relevance feedback, i.e. by allowing the user to progressively refine the system's response to his query. The key concept of the relevance feedback we propose is the statistical analysis of the feature distributions of the retrieved images the user has judged relevant, or not relevant, in order to understand what features he has taken into account (and to what extent) in formulating this judgment, so that we can then accentuate their influence in the overall evaluation of image similarity, as well as in the formulation of a new query iteration.

We have addressed the problem of retrieving perceptually “similar” images in a database of trademarks containing a great variety of objects. The performance of an image retrieval system is closely related to the nature and quality of the features used to represent image content, but it is also strongly influenced by the measure adopted to quantify image similarity. We have shown that the use of relevance feedback greatly improves retrieval results, making it possible, in many cases, and with no significant effort by the user, to tune the similarity measure used by the system to the user's notion of image similarity. Designing the whole system we have assumed that the user will identify at least one relevant example in the database by random browsing and, that some relevant images will be retrieved at the first iteration (in which all the features have the same importance) within the first set of displayed images. In our experiments we found that these conditions were almost always met for the database used, but we realized that this might not be the case if we scale up the size of the database. Finding an initial query image will be further addressed in our system by allowing the user to perform the query with a sketch, or using an image from outside the database, and by offering a database preview not only by random access, but also by image clustering. The lack of relevant images retrieved at the first iteration, could be addressed pragmatically by allowing the user to select relevant and not-relevant images not only within the first 24 retrieved images (we found that in many cases relevant images were ranked within the first two of three screens). For faster tuning of the similarity function, which would also deal in part with this problem, we could also exploit previous query sessions performed by the user on the same database. The user would be allowed to register satisfactory queries, together with the corresponding weights in the similarity measure. When the user has already formulated a query “similar” to the new one, the algorithm could set the initial weights of the similarity function at the value of the earlier query, reducing the time and effort needed to adapt the similarity measure by applying the relevance feedback algorithm.