Target With Variability Recognition Based On Local Invariant Feature

Along with target recognition in an image,variability of the image often happens,which means the angle of view,the light and the scale in the image often changes.This is one of the most difficult part of image recognition.Local feature is an alternative way to solve this problem since it has good robustness.In this paper,a local feature based image recognition algorithm is proposed to improve the real-time performance of the recognition,assuring better robustness.The target recognition in this paper mainly includes three parts: feature extraction,feature coding and feature classification.The Feature extraction if the most important part of target recognition algorithm,and it is also the core of this paper.The multi-scale space and multiple kinds of key point detector are introduced and some classical algorithms are compared.Each algorithm's advantages and disadvantages are pointed out.Based on the classical algorithms,this paper proposes an algorithm combined with improved FAST key point detectors based on machine learning and improved SURF descriptor.The experiment shows that this algorithm has better real-time performance and better robustness.Bag of Features is introduced in feature coding part.In this paper,the Bag of Features coding algorithm based on spatial information is used to get higher performance during the classification part.At last,we use Gaussian Kernel-based Support Vector Machine to classify the feature vectors,and the whole target recognition is completed here.We use database of ALOI to do the experiment.The result shows that the algorithm this paper proposed achieves 98.8% and 93.4% recognition rate when viewpoint and light are changing.The average time cost is 109.0ms and 118.2ms.Also,this algorithm has better performance when scale changes.