Reaseach On Recognition Tech Of Armyman Uniform Pictures Based On Multiple Features Fusion

With the development of information society, soldiers have more and more chancesto access to Internet, soldiers’ accesses to Internet have become more and more frequent.Some active-service and ex-service soldiers liked to upload theirselves’ uniform photosto some forums and social network sites(such as QQ, renren, etc.), which have broughtnew challenges to protect army’s secrecy and image. Hence, a series of militaryregulations have been issued in order to regulate soldiers’ Internet behavior. Forenhancing detection abilities of soldiers’ irregular behavior, in this dissertation, theproblem of armyman picture recognition was researched and in which a new recognitionmethod based on fusing multiple features was proposed, this method can be used todetect soldiers’ Internet behavior effectively. Firstly the features of color, texture andlocal areas were gotten from upper body of the picture, and then the features were fusedand the pictures were classified by multiple kernel learning method.The main contributions in this dissertation are summarized as follows:Firstly, a new image feature representation method named fused SIFT-BOW basedon fusing sparse SIFT and dense SIFT is proposed. Local salient regions in an imagedetected by SIFT algorithm are described by sparse SIFT as foreground, the rest regionsin the image are described by dense SIFT as background. To generate fused SIFT-BOW,the sparse SIFT and dense SIFT are combined with adaptive weights, as a result, sparseSIFT is enhanced as main descriptor and dense SIFT is weakened as secondarydescriptor. The fused SIFT-BOW method combines the advantages of sparse SIFT anddense SIFT, and can describes local silent areas and flat areas of armyman uniformmore accurately.Secondly, a new BOW model ESWM-BOW which is based on new weightingmethod named Entropy Spatial Weight Map (ESWM) is proposed to abstract thefeatures of local areas of armyman uniform picture. Spatial Weight Map (SWM) iscomputed according to the frequency of visual word appeared in different positions.Because the probability of a visual word appears in different image categories may bedifferent, entropy in information theory is used to describe the probability. In order tocompute ESWM, not only SWM, but also the entropy of visual word which representscategory information is considered. ESWM can improve the discrimination performanceof visual words. Thirdly, multiple kernel learning is adopted to fuse multiple features to improve theperformance of classifier. The fittest kernel function is chosen for each kind of feature,and then the weight of kernel function is optimized by multiple kernel learning.Compared with support vector machine based on single kernel function, support vectormachine based on multiple kernel learning can gain better classification performance.The new method of armyman image recognition is simulated by matlab7.10,experiment results reveal that this method can recognize most armyman uniformpictures effectively, and be robust for illumination, partial occlusion and posture change.The average accuracy of classification can reach up to92.8%.