Research On Image Feature Extraction Method For Design Patent Image Retrieval

With the improvement of the intelligent properties laws and the enhancement ofhuman legality awareness, the number of applying patents is increasing dramatically fromindividuals and enterprise. Currently in domestic, the design patents retrieval systems aremainly based on text description, but those systems could not satisfy users’ needs.Meanwhile Content-Based Image Retrieval (CBIR) systems are usually for generalpurpose, which works on images from Internet, So general CBIR systems are not practicalfor patents image retrieval.By analysing the attributes of patent images, the low-level visual features and highsemantic features are used for content-based patent image retrieval system. The researchfocus on image feature extracting method and discriminative ability. For the reasons above,this thesis focus on how to implement design patents images retrieval. The low-levelvisual features contains colour and shape features. Colour feature is obtained by dividingthe images into small grids and then get colour histogram in grids. Shape feature isobtained by pyramid of histogram of orientation gradients. Spatial information isintegrated into global features for closing to users’ perception. Moreover, the knowledgeof images used for applying patents is defined quite concrete and specific, which is goodfor aliasing semantic to images. By taking advantage of design patent images’ semantics,in addition to low-level visual features, semantic features of patent images are used. Forthese special properties, a framework is proposed for patents images semantic featuresextraction. In implementing semantic features, semantic space is defined by selectingsample images and the space reference basis is acquired by the hyper-plane derived fromsupport vector machine. The low-level visual features under semantic feature areintegration of appearance and geometry shape features. Appearance features has takenadvantage of local features, but avoided its key-points detecting part which is quitetime-consuming, so this feature could meet the requirement of real-time applications.Images features are evaluated in discriminative ability, feature’s dimension andextraction time speed on testing images from Chinese design patent. The testing resultsdemonstrated that these features can meet the requirements of retrieval on speed andperformance. What’s more, the semantic feature outweighs other features significantlyincluding the state-of-art local features SIFT in discriminative ability.