Research On The Feature Description And Similarity Measure Of NAM Images Based On Keypoints

With the popular use of the devices on image acquirment and data transmission image data is dramatically increasing. Image analysis and understanding become the important issue of content-based multimedia data management. With respect to object classification in image comparing, based on NAM (Non-symmetry and Anti-packing Model) and SIFT (Scale Invariant Feature Transform), a novel grey image feature extraction, representation, and description method is proposed, the similarity retrieval performance of the feature sets is studied. The new approach gives more efficient and effective on image classification by experiment.Firstly, large scale areas in the image are extracted fastly. Research on visual psychology showed global object dominance during the beginning of visual perception. Global objects in images could be regarded as the large scale regionals with smooth gratation. In order to abstract the global object in image efficiently, keypoints of SIFT are adapted to the start points of NAM subpatterns, the gratation smooth rectangle subpatterns are extended from keypoints of SIFT by a range of grey value, as it combined SIFT and NAM, the approach is called as SNAM. The feature sets is invariant for size and radiation deformation. The additional computation cost emploied than SIFT is O(M) and M is the size of a image.Secondly, SNAM descriptors significantly outperformed SIFT descriptors for image feature representation. In SNAM, area vector, position vector, and location vector are defined as to represent the area, positional, and local features of global objects. The area vector and the positon vector are single variable in 16 dimention, and the location vector is two variables in 16 dimention. But, in SIFT, single variable in 128 dimention was assigned to each keypoint. SNAM descriptors are more effective to represent the higher feature of images, such as the position, distribution, and location features of the large scale areas. A set of experiments demonstrated that the newly method is available for distinguishing and clustering images by the object contents. The computation cost of the three SNAM descreptors are respectively O(n), O(n), and O(n2), n is the number of subpatterns.Thirdly, the pecision of siminarity retrieval was increased by SNAM. The siminarity retrieval performance of the area vector, position vector, and location vector are detailed. Ordered distance, nominated distance, and Eulicdean distance are adorpted during similarity measurement. Experimental results show that this approach performs better than conventional approaches in recall rate and pecision at a less retrieval results. The algorithm of similarity measure has an time complexity of O(b2), and b is the dimention of the vector. As the feature descriptor has less size of data, the algorithm has lower space complexity.Finally, the parameters of SVM kenel classifiers using area vector and position vector as input descriptors were optimated. Experiments on area vector and position vector with SVM kenel optimation parameters denote a steady classification accuarcy at 75%, the samples for learning is 60% and others for testing. The effective kernels are inner product, polynomial, and Gaussian, for polynomianl kernels, degrees below 5, for Gaussian kernels,σvalues below 0.1.According to the theory of visual perception, combine NAM and SIFT, A new region feature description is proposed, and 3 feature vectors are defined in 16 dimentions. Experimental resuls show the approach performs rapidity retrival in similarity measure and image classification.