FEATURE SPACE DESIGN FOR STATISTICAL IMAGE RECOGNITION WITH IMAGE SCREENING

This paper proposes a design method of feature spaces in a two-stage i mage recognition method that improves the recognition accuracy and eff iciency in statistical image recognition. The two stages are (1) image screening and (2) image recognition. Statistical image recognition me thods require a lot of calculations for spatially matching between sub images and reference patterns of the specified objects to be detected in input images. Our image screening method is effective in lowering t he calculation load and improving recognition accuracy. This method se lects a candidate set of subimages similar to those in the object clas s by using a lower dimensional feature vector, while rejecting the res t. Since a set of selected subimages is recognized by using a higher d imensional feature vector, overall recognition efficiency is improved. The classifier for recognition is designed from the selected subimage s and also improves recognition accuracy, since the selected subimages are less contaminated than the originals. Even when conventional reco gnition methods based on linear transformation algorithms, i.e. princi pal component analysis (PCA) and projection pursuit (PP), are applied to the recognition stage in our method, recognition accuracy and effic iency may be improved. A new criterion, called a screening criterion, for measuring overall efficiency and accuracy of image recognition is introduced to efficiently design the feature spaces of image screening and recognition. The feature space for image screening are empiricall y designed subject to taking the lower number of dimensions for the fe ature space referred to as L-S and the larger value of the screening c riterion. Then, the recognition feature space which number of dimensio ns is referred to as L-R is designed under the condition L-S less than or equal to L-R. The two detection tasks were conducted in order to e xamine the performance of image screening. One task is to detect the e ye-and-mouth-areas in a face image and the other is to detect the text -area in a document image. The experimental results demonstrate that i mage screening for these two tasks improves both recognition accuracy and throughput when compared to the conventional one-stage recognition method.