Case Study An Empirical Investigation of Thumbnail Image Recognition C. A. Burton, L. J. Johnston and E. A. Sonenberg The University of Melbourne, Parkville, Victoria, Australia, 3052. Abstract and noise and background texture smoothed [14]. The use of thumbnails (i.e. miniatures) in the user-interface of image databases allows searching and selection of images without the need for naming poli- cies. Treating parent images prior to reduction with edge-detecting smoothing, lossy image compression, or static codebook compression resulted in thumbnails where the distortion caused by reduction was lessened. An experiment assessing these techniques found result- ing thumbnails could be recognised more quickly and accurately than thumbnails of the same parent images that had been reduced without treatment. This pre- treatment in thumbnail creation is offered as an im- provement for browsing of image databases. 1 Introduction The increased popularity of image databases has encouraged user interface designers to use highly re- duced images or thumbnails to facilitate the handling of images. Thumbnails are displayed in place of full- sized images to allow quick, preliminary visual iden- tification. Such reduced images require much less bandwidth for transmission; they are very common in World Wide Web pages. In databases, thumbnail use overcomes the difficult process of naming images, frees image identification from problems of language and literacy, and encourages browsing and manual or- ganisation by related visual content. Another approach is to use lossy image compres- sion such as the Joint Photographic Experts Group (JPEG) method. The objective of modern lossy im- age compression techniques is to reduce the file size while maintaining important perceptual information. Extreme compression of an image will in effect sim- plify an image by removing many of the colours and textures. By taking the original image closer to an abstract iconographic form, it may be expected that such compression would suit thumbnail creation. In a perceptual study of icon designs for user interfaces, icons that were more abstract and visually simpler were identified as being easier to understand than complex images or complicated symbols [3]. It is pos- tulated that “simplified” thumbnail images will en- hance a user’s ability to browse, as perception has been shown to play an important role in browsing [S]. Because thumbnail images are such dramatic re- ductions of their parents (for example a 64x64 pixel thumbnail from a 1024x1024 parent) distortions, such as exaggerated textures and stepped edges, occur dur- ing the subsampling used to achieve reduction. This article investigates whether thumbnail creation could benefit from image-enhancement technology being ap- plied before subsampling. In the work described below, EDS, JPEG and a self- organising feature map (SOFM) were used to treat full-size images as an enhancement step before the usual subsampling that reduces their physical size. Forty-eight computer users viewed one hundred and sixty of these thumbnails in an experiment designed to explore the value of such thumbnail creation tech- niques in a context akin to browsing, where the viewer has some idea of what they are seeking. Section 2 presents the issues central to the selection of pre- treatment processes. The design of the experiment and some key results are provided in Section 3. Sec- tion 4 includes our discussion and conclusions. 2 Image Simplification One approach is to apply edge-detecting smoothing (EDS) which enhances images by smoothing around, but not over, large changes in image intensity, result- ing in visually simpler images with edges preserved In evaluating methods for removal of information from an image while retaining fast accurate recog- nition, one must consider: factors influencing object recognition; how to effect improvements on thumbnail images which might lead to better recognition per- formance; and the choice of measures to determine thumbnail quality. O-8186-7201-3/95 $04.00 0 1995 IEEE 115 See Color Plates, page 150. Proceedings of the Proceedings on Information Visualization (INFOVIS '95) 0-8186-7201-3/95 $10.00 © 1995 IEEE Authorized licensed use limited to: Monash University. Downloaded on February 05,2023 at 08:15:57 UTC from IEEE Xplore. Restrictions apply.