An Algorithm For Automatic Target Recognition Using Local Features

The difficulty of automatic long-distance target recognition lies in the backgroundcluttered and relatively less pixel the targets have, which factors make long-distancetargets difficult to describe. This dissertation proposed a recognition algorithm usinglocal features against long-distance targets, its procedure includes local featureextraction, local feature encoding and feature classification.First, several of the typical key-point detectors are compared, and the statement ismade that detectors with Gaussian blur are apt to filter out some of the necessarydetails and are thus disadvantageous to long-distance target description. Scale-spaceAGAST corner detector is therefore used. A mathematical derivation is also made toimprove the SURF descriptor, and the improvement has gained the descriptor a betterresistance to image rotation. Based on these studies, the proposed algorithm thereforeextracts local features using scale-space AGAST detector and the improved SURFdescriptor.The local features are then encoded using the Bags of Features algorithm basedon Spatial Pyramid Matching. The encoded features characterize the targets moreeffectively.Finally, Support Vector Machine with Radial Basis Function kernel is used forfeature classification, allowing for a millisecond-level performance. The experimental results have shown that the proposed algorithm has reachedhigher computation efficiency (163.9ms) than its counterpart using the original SURFfeature (213.4ms); a relatively high recognition rate against targets with view-pointchange (96.61%) or illumination change (96.88%); and a recognition rate of above50%against targets with scale change which are obtained via image down-samplingby a factor of7, far exceeding that of the algorithms using either SIFT or SURF.An analysis over the results is made, and the conclusion can be drawn that theproposed algorithm is well applicable to long-distance target recognition.