Research On Algorithms In Image Processing And Object Recognition Based On Graph Theory

Object recognition is a challenge subject in the field of computer vision. The challenges lie in three processes of the object recognition system: image acquisition, image low-level processing, and object recognition. It is not uncommon that critical information can be lost during the image acquisition process. During the subsequent image low-level processing, clear regions of an object in an image may not be obtained. Lastly, it is challenging to recognize generic objects in images taken under different imaging conditions based on characteristics descriptions. Therefore, the algorithms relative to image acquisition, image processing, and object recognition warrant further study and this paper presents a series of algorithms to enhance the capability of the object recognition system.To acquire clear images without critical information loss, an adaptive image fusion algorithm based on algebraic multigrid method is proposed. The algorithm helps to extract coarse-level details for the reconstruction of the original images. Our experiments have showed that the algorithm is promising in image reconstruction and theoretical analysis is performed in this study to further substantiate the effectiveness of the algebraic multigrid method in extracting the strongly connected subgraphs. The mean square error (MSE) between the reconstructed image and the original image is analyzed. When the original image is clear, the MSE is large and when the image is blur, the MSE is small. This principle is used in this study to guide the fusion of multi-focus images. The fusion results show that this method results in better results over other methods based on subjective and objective evaluation criteria.To obtain the information of the shapes and regions, low-level processing including image denoising and image segmentation is performed. K-means median filtering algorithm and recursive K-means median filtering algorithm are proposed to denoise the image. These approaches can eliminate some errors caused by the standard median filtering algorithm and greatly optimize the time-consuming K-means algorithm. The relationship between the graph cut method and the algebraic multigrid method is analyzed and a method to extract the coarse level details is proposed based on the eigenvalues of the laplacian matrix of the adjacency graph. The results are consistent with two basic principles of coarse grid selection. A new concept of "wavelet transformation of the graph" is proposed to extend the application of the graph in image processing.For the graph classification methods based on the energy function, we have established an energy function equivalent to energy functions based on regional growth method, K-means method and other methods. For the specific image segmentation, a variety of features combined with normalized cut, such as wavelet operator, fractional differential operator, and algebraic multi-grid operator are presented. These operators are capable of obtaining finer structure and texture features. The OTSU method is combined with normalized cut method in this study to get more accurate contour feature.In the recognition of single kind of object, promising results can be achieved after learning and training through the support vector machine method combined with appropriate image features and characteristics descriptions. In the recognition of different kind of objects, an adaptive "word bag" image expression combined with region segmentation and graph theory is proposed. According to the relationships of regional characteristics and some other characteristics, the prominent characteristics are strengthened and others are inhibited.