Content Based Trademark Retrieval Using Scale Invariant Feature Transform

Abstract

Trademarks play very important role for successful companies to establish their brand name in the market. Trademark registration and evaluation has become tedious job for registration officers because of millions of trademarks already registered and it is very difficult to compare new symbol with existing trademarks to ensure the similarity between them. Trademark Image Registration is an important application area of Content Based Image Retrieval (CBIR). CBIR is the process of finding similar images from large image databases which is a challenging task. This is a research area that having practical importance when applied to Trademark Image Registration. Common approaches use only low level features to match images. Such CBIR solutions fail on capturing some local features representing the details of images. Local image descriptors have received much attention in this regard because of their efficiency and accuracy. Among them Scale Invariant Feature Transform (SIFT) vectors have been widely used in lot of applications and identified as the most efficient local feature description algorithm based on scale-space. This research attempts to propose a novel method for trademark image retrieval. Proposed system will combine global feature extraction method with local feature extraction algorithm in order to improve the performance. As the global feature extraction method it will use the color based image retrieval approach and SIFT algorithm will use as the local feature extraction method. First, the 3D (HSV) histograms of all images in the database are computed. Based on that, the system retrieves subset of visually similar images to the query image. SIFT matching is performed over the retrieved results to retrieve most relevant results. Results have been evaluated on database of hundred trademark images belonging to twenty different categories with different orientations, scales and illumination levels. System prototype has been tested with two bogus trademarks from each category. Experiments have been conducted to show that the developed method is useful to retrieve similar symbol from the collection of trademark images and retrieval time is significantly less than the time taken by the SIFT alone. The proposed method gives 50% of retrieval efficiency on tested trademark image database.