Study On Image Object Recognition Method Based On Complex Network

Image Object Recognition technology is one of the most important areas of Artificial Intelligence research. It had lots of successful applications such as Video Monitoring, Human-Computer Interaction, Traffic Monitoring, Behavior Recognition, Automatic Navigation, etc. There are several valuable methods for object recognition. According to the features of target object which are used in the recognition process, the object recognition methods can be roughly divided into four categories, such as the Region-Based algorithm, the Contour-Based algorithm, the Model-Based algorithm and the Feature-Based algorithm. The algorithms which are mainly used in object recognition methods can be divided into the Filtering theory, the Mean Shift method, the Partial Differential Equations method, etc. It should be noted that these recognition methods are closely related to the location and sequence of the pixels in the image. So the recognition rate will be reduced by rotation, translational and scaling. In the practical application, the slight changes of the image contour, small variations of the light intensity as well as partial occlusion and other factors will produce adverse effects on the recognition results.Getting more and more attention by the researchers, Complex Network theory becomes important in the research of the Complex System. The concepts and methods of it have become research focus recently. Modeling by the mathematical Graph theory,Complex Network theory considers only the topological relationships, such as the relative position between nodes. It pays less attention on the location and sequence of the nodes. The overall rotation and translation will not affect the topology of the network diagram. So, a shape recognition algorithm, which uses the Complex Network theory as the description model to the shape of image boundary, may be able to effectively adapt to the changes in the image boundary.In this work, the Complex Network theory was used in the Image Object Recognition. Based on the existing researches, an image object recognition method based on Complex Network was applied to the shape contour recognition and gray image recognition applications. The proposed method could have the advantages of both Contour-Based recognition method and Complex Network theory, including simple process, high identify efficiency and regardless of the position or order of the nodes in the image. The proposed method can overcome the slight changes of the image contours and the light intensity as well as partial occlusion. In this way, the recognition method is rotation invariance, translation invariance,scale invariance and noise tolerance.This work takes the Image Object Recognition method in the two-dimensional image sequences into account. By extracting contours from the object, the proposed method recognizes the object from the image sequences. This work will provide reliable data for target tracking, behavior analysis, description and understanding. The main technical route of the proposed method could be described as follow. Firstly, the shape contour and color contours are extracted separately from the image object. Then, the contours are described into graphs and modeled by the Complex Network methodology. Finally, the characteristic parameters are calculated for each network model. And a feature vector for object recognition is extracted for image object recognition and classification.This work mainly includes the following aspects:(1) Combining the Complex Network theory and Contour Recognition method together. This work extracts the target identification parameters by the topology information of the contour with Complex Network theory. In this way, the method has the advantages of the Complex Network theory and the Contour Recognition method. With a simpler network model, the proposed method becomes noise tolerance and reaches effective recognition result.(2) Controlling the complexity of the network. In this work, several methods are used to control the size of the complex network model. Firstly, the proposed method reduces the number of nodes in the network model by extracting the contours from the image. Secondly, the proposed method uses the simpler network parameters in order to reducing the complexity of calculating. Thirdly, the process of modeling is improved. By these changes, the storage space occupied by the recognition method is reduced and the calculation time is shortened. Topology information, as much as possible, is extracted from these simpler network models for object recognition.(3) Improving the image contour extraction method. In this work, a set of shape contour and color contours are extracted from each grayscale image for object recognition. Taking use of the shape information and the color information, the proposed method enhances the differences between recognition targets and improves the recognition efficiency. Among these, a color contour extracting method was proposed by removing the adjacent points from the binary image. It can keep the color information from the image while reducing the pixels from the binary image. (4) Adjusting the threshold parameters and identify parameters. In this work, new methods are proposed to help choosing the threshold parameters and identify parameters. A Distance Threshold Determing (DTD) method is used in determining the distance thresholds. By calculating the determining parameters, the DTD method could reduce the scope of values for distance threshold choosing. Meanwhile, a set of network parameters based on node degree is used as recognition parameters. In this condition, more topological characteristics of the network model are extracted for recognition. The complexity of calculating is also reduced.Experiments show that the proposed method could effectively control the scale of the complex network and adapt to the changes in the boundary shape with efficient power of object recognition. The data from the experiments shows that the proposed method is scale invariant and rotation invariant. The DTD method is also proved to be effective.This work was supported by the National Program Foundation (61273219), National Natural Science Foundation of Guangdong Province (No.8151009001000061), and Natural Science Joint Research Program Foundation of Guangdong Province (No.8351009001000002).