Image Based Design Patent Retrieval with Classification and Indexing

As a new image retrieval technology, Content Based Image Retrieval (CBIR) has a broad application prospect. In a normal CBIR system, the mature technology is usually as follows.Firstly, the underlying features of the image in the database are extracted in the preprocessing process. Then, during the retrieval process, the feature is matched with the feature of the Query Image linearly [1], [2], [3]. As the existence of semantic gap, this technique does not perform well enough to satisfy users’ need.Compared with the general image on the network, the design patent images have some special properties. For example, the object in one image is singular, the images are fine categorized and little blurred, etc (e.g. Fig. 1). Based on these good properties, this paper presents an image retrieval method of design patent based on classification technology. When the retrieved image is correctly classified with a high probability, it is retrieved in the category; otherwise, it is retrieved in all images by using indexing technology to speed up the retrieval.

I. Introduction

The main aim of this paper is to retrieve similar design patent images based on image content accurately and rapidly for users such as patent agents and design patent inventors. These users need to check whether a patent is innovative or not. Design patent images which present the main point of the design patent are the most important part of this patent (e.g. Fig. 1). Several web sites provide patent retrieval service now in China [4], [5]. But almost all of these retrieval systems just provide text based patent retrieval. The text-based retrieval service can not meet the user's retrieval demand for similar design patent images. As a result, the search of design patents based on image content is becoming more and more important to patent attorney and inventors of design patents.