Multi-feature Fusion Research And Implementation Applied For Target Recognition

The history of research on object recognition has been more than fifty years.Prior to 2012,target recognition model included manual features and classifier,usually.After 2012,the CNN(Convolutional Neural Networks)model which can be trained end-to-end has become a research hotspot.Due to the high precision of deep learning,some scholars try to use the features of deep learning to replace the traditional features in various aspects.Considering the generalization performance of single model and the poor interpretability of deep learning itself,some other scholars are exploring various fusion schemes.At present,it is generally accepted in academia that fusion schemes are classified into three levels:data-level,feature-level,and decision-making level.Data-level fusion scheme is used in multi-sensor information fusion.Due to the limitations of computer,earlier studies mainly focused on the fusion of decision-making level.In recent years,with the improvement of computer performance,feature-level fusion has become a research hotspot.At present,we are still exploring the fusion scheme of a certain combination of traditional features and the performance of the fusion scheme in a particular task.With the development of CNN,it is worth exploring to combine many traditional features including combining traditional and deep learning features.This article implements three kinds of feature fusion methods:(1)Feature fusion among traditional features.We choose different combinations of serial fusion in five typical features:color histogram,LBP(Local Binary Pattern)histogram,HOG(Histogram of Oriented Gradient),mean values of Gabor filter response,BoW(Bag of words)based on dense SIFT(Scale-invariant feature transform),and followed by a fully connected neural network.(2)Traditional Features fuse with deep learning features directly.Five typical features are selected:color histogram,LBP statistical histogram,HOG,mean values of Gabor filter response,BoW based on dense SIFT,and trained features of CNN.(3)Traditional features fuse with deep learning feature indirectly.Four typical features are selected:Canny edge feature,LBP texture feature,Gabor texture feature,dense SIFT feature.Traditional features are indirectly combined with deep learning features in different ways.The fusion results are tested on the CifarlO standard database.Compared with a single feature,the fusion of the traditional features is better than a single feature,but it is still lower than the deep learning feature.The fusion of traditional features and deep learning features in direct and indirect ways are both better than single deep learning.Improve the performance of the traditional SPM(Spatial Pyramid Matching)algorithm.The SLIC(simple linear iterative cluster)algorithm is used to improve the SPM model based on dense SIFT.Finally,we collect and establish the Summer Palace database,using which we train a model.We apply the model in Android platform to achieve tourist guide based on image recognition algorithm.