The Research On Object Recognition Based On The Multi-object Relationship

With the rapid development of computer technology,image and video data has reached aspects of daily life.What's more,the massive application of object recognition promotes the position in different fields.Meanwhile,the rapid growth of image and video data has promoted more research in computer vision,pattern recognition,artificial intelligence and other related fields.Object recognition which plays as an important branch in computer vision has been applied in aerospace field,military field,public security field and related industrial and agriculture filed.There is no doubt that it has widespread value and prospect for application.However,object recognition remains a challenge research problem due to the limitation of feature extraction and recognition method.This paper studies an object recognition method based on multi-object relationship,whose aim is to recognize objects in image through utilizing multi-object relationship among objects.Given an image,we first run object detectors to output a set of candidate detector windows.Each detector window includes category information,location information and score information.Then we apply high-order dependence and Bayes rule to obtain semantic contexts,spatial and scale contexts among objects in detector windows.Moreover,we construct a graph of objects' relationship based it.Meanwhile,semantic contexts,spatial and scale contexts belong to local contexts.Then this paper implements a BLSTM-RNNs model to obtain the global context of each image.All candidate detector windows and Gist feature information of an image are as input respectively.This article adopts a bidirectional long short-term memory recurrent neural network(BLSTM-RNNs)to deal with the problem of variable-length sequence produced by outputs of object detectors in different images.Afterwards,we train BLSTM-RNNs to generate the hidden state corresponding to each detector window and fuse each forward hidden state and backward hidden state corresponding to each detector window.After getting all the forward hidden states and backward hidden states,this model calculates their average and uses it as the global context feature of the detector window sequence.Meanwhile,the obtained feature and the Gist feature information are fused.Then it can be used as the global context feature of the given image.Genetic algorithm is used for searching for the optimal solution in the framework of object recognition which corresponding to the most likely correct windows combination we aim to find.The corresponding fitness function is constructed by the fitness of singe object,the fitness of pairwise objects connected by edge and multi objects enclosed by cliques.And the object in the graph of objects' relationship is correspond to the gene in the chromosome.Moreover,the fitness of single object is determined by location information of the detector window,score information of the detector window and the obtained global context of an image.The others are decided by the constructed graph.After initializing all the population in random way,we use the obtained fitness function to judge the fitness of the chromosome.After many generations,we can obtain the optimal solution which corresponding to the most likely correct windows combination.In experiment part,the method based on multi-object relationship is validated in SUN09 dataset.It includes rich relationship among objects.The experiments results illustrate that the proposed method boosts the performance of object recognition,and it has potential application value in object recognition.