Research On Object Image Recognition Under Complex Background

Judging from the development of word military situation, using accuracy guidance weapon to form conventional deterrence force to capture war initiative is one mode of the future war. With the development of optical current technology, image guidance is becoming one of the important methods for accuracy guidance technology. As object recognition is one of the key technologies in image guidance and the battlefield environment is always complex, to research High-precision object recognition algorithm is very important. So the paper researched the object recognition under complex background, including image preprocess, image feature extraction, object recognition and multiple classifiers combination.Firstly, according to the aim of project, the paper discussed the image preprocess under complex background. In view of characteristic of the tank image: this paper segment the object from complex background using median filtering, threshold segmentation, mathematical morphology, edge detecting, contour trace and region filling .Secondly, it extracted shape features to descript the attributes of object. After eliminating the redundancy in features it selected Form factor, Eccentricity, Sphericity, Circularity, Rectangularity and Zernike moment as the object recognition features to provide efficacious data for object recognition.On the basis of the research above-mentioned, it deeply studied the object recognition technology based on BP neural network, designed a classifier based on BP neural network. The design steps including determination the nodes number of input and output of BP neural network, selecting the structure of network (nodes number of hidden layers) and the training parameters of network.In the end, considering single classifier is difficult to obtain satisfactory results, the paper introduced an improved multiple classifiers combination algorithm based on voting algorithm which regarded selected features as inputs and designed three sub networks: geometry features network, conventional moment network and Zernike moment network which were based on BP neural network. Then it used two-stage voting algorithm to combine sub network classifiers to enhance the reliability of object recognition.