Research On Multi-view Indoor Object Recognition Algorithm

Stationary object recognition is a part of the field of computer vision. Shooting angle ofthe camera have a great effect on the external characteristics of the object. The generaldirection of the present study is divided into two major aspects. One is dedicated todetection or classification. The other is the direction of image matching. A finite number ofobjects’ images constitute image database. After obtaining the actual image, the systemmatches it with the images in the database. Then it outputs the input image’s categorycorresponding to the image database. This paper focuses on the latter and discusses somework on this subject. The following will describe the main contents.This paper studies on the target extraction. It puts forward a algorithm of targetextraction. It uses the color histogram combined with threshold to get the initial target areas.Then it extracts continuous edges from the gray image. Through contour matching and theexclusive or operation, it finds the distinct regions. Finally, to get the final object areas, itcalculates the area distance to modify the initial target areas. Using the color information, itcan quickly extract out the rough regions of the targets to save the operation time. Colorfeature combining with spatial feature avoids the over-segmentation or under-segmentationphenomenon.This paper researches on the object feature extraction. It constructs a comprehensivecharacteristics extraction method. In terms of shape feature extraction, this algorithmselects the top five Hu moments. Because Hu moments can’t reflect the affine information,it extracts six affine invariant moment features to complement the shape information ofobjects. In the color feature extraction, it extracts six color moments. In the texture featureextraction, it extracts the means and standard deviations of the energy, entropy, moment ofinertia and correlation. This combination of these eight parameters describes the textureinformation.This paper also researches on the object recognition. The experiment selected7200images of100classes objects from COIL-100image database and1296images of18classes object which were taken by myself. It extracted Hu invariant moments, colorinvariant moments, affine invariant moments, texture features and integrated features. Bysimulation experiments using Support Vector Machine (SVM), extreme learning machine,nearest neighbor classifier methods, it proved that support vector machine objectrecognition method based on the comprehensive feature has high recognition accuracy andgood stability.